Downloaded 25 times
![M
a
r
c
h
2
6
,
2
0
2
4
/
D
r
.
R
S
processing system.
– Stream Processing Engine:— This is the core component of the stream processing system,
which processes the incoming stream data. The processing engine is responsible for ingesting
data2
, applying transformations, aggregating data, and producing output data in real-time.
Output Sinks:— These are the destinations of the processed stream data. The output sinks
may include databases, message queues, dashboards, or other systems that consume the output
data.
Figure 1: Illustration of Architecture of Data Stream Model [16].
The architecture (See Figure 1) of a stream processing system can be implemented using a variety of
technologies, such as message queues, stream processing frameworks, or microservices.
ˆ Stream Processing Frameworks—Stream processing frameworks are a popular choice for imple-
menting stream processing systems. These frameworks provide a set of libraries, APIs, and tools for
2Data ingestion is the process of importing large, assorted data files from multiple sources into a single, cloud-based storage
medium—a data warehouse, data mart or database—where it can be accessed and analyzed.
5](https://image.slidesharecdn.com/kit-601lecturenotes-unit-3-240523065037-f5d66706/75/KIT-601-Lecture-Notes-UNIT-3-pdf-Mining-Data-Stream-5-2048.jpg)
![4.9. REFERENCES FOR CHAPTER 4 161
✦ Estimating the Number of 1’s in a Window: We can estimate the number
of 1’s in a window of 0’s and 1’s by grouping the 1’s into buckets. Each
bucket has a number of 1’s that is a power of 2; there are one or two
buckets of each size, and sizes never decrease as we go back in time. If
we record only the position and size of the buckets, we can represent the
contents of a window of size N with O(log2
N) space.
✦ Answering Queries About Numbers of 1’s: If we want to know the approx-
imate numbers of 1’s in the most recent k elements of a binary stream,
we find the earliest bucket B that is at least partially within the last k
positions of the window and estimate the number of 1’s to be the sum of
the sizes of each of the more recent buckets plus half the size of B. This
estimate can never be off by more that 50% of the true count of 1’s.
✦ Closer Approximations to the Number of 1’s: By changing the rule for
how many buckets of a given size can exist in the representation of a
binary window, so that either r or r − 1 of a given size may exist, we can
assure that the approximation to the true number of 1’s is never off by
more than 1/r.
✦ Exponentially Decaying Windows: Rather than fixing a window size, we
can imagine that the window consists of all the elements that ever arrived
in the stream, but with the element that arrived t time units ago weighted
by e−ct
for some time-constant c. Doing so allows us to maintain certain
summaries of an exponentially decaying window easily. For instance, the
weighted sum of elements can be recomputed, when a new element arrives,
by multiplying the old sum by 1 − c and then adding the new element.
✦ Maintaining Frequent Elements in an Exponentially Decaying Window:
We can imagine that each item is represented by a binary stream, where
0 means the item was not the element arriving at a given time, and 1
means that it was. We can find the elements whose sum of their binary
stream is at least 1/2. When a new element arrives, multiply all recorded
sums by 1 minus the time constant, add 1 to the count of the item that
just arrived, and delete from the record any item whose sum has fallen
below 1/2.
4.9 References for Chapter 4
Many ideas associated with stream management appear in the “chronicle data
model” of [8]. An early survey of research in stream-management systems is
[2]. Also, [6] is a recent book on the subject of stream management.
The sampling technique of Section 4.2 is from [7]. The Bloom Filter is
generally attributed to [3], although essentially the same technique appeared as
“superimposed codes” in [9].
M
a
r
c
h
2
6
,
2
0
2
4
/
D
r
.
R
S](https://image.slidesharecdn.com/kit-601lecturenotes-unit-3-240523065037-f5d66706/75/KIT-601-Lecture-Notes-UNIT-3-pdf-Mining-Data-Stream-47-2048.jpg)
![162 CHAPTER 4. MINING DATA STREAMS
The algorithm for counting distinct elements is essentially that of [5], al-
though the particular method we described appears in [1]. The latter is also
the source for the algorithm for calculating the surprise number and higher
moments. However, the technique for maintaining a uniformly chosen sample
of positions in the stream is called “reservoir sampling” and comes from [10].
The technique for approximately counting 1’s in a window is from [4].
1. N. Alon, Y. Matias, and M. Szegedy, “The space complexity of approxi-
mating frequency moments,” 28th ACM Symposium on Theory of Com-
puting, pp. 20–29, 1996.
2. B. Babcock, S. Babu, M. Datar, R. Motwani, and J. Widom, “Models
and issues in data stream systems,” Symposium on Principles of Database
Systems, pp. 1–16, 2002.
3. B.H. Bloom, “Space/time trade-offs in hash coding with allowable errors,”
Comm. ACM 13:7, pp. 422–426, 1970.
4. M. Datar, A. Gionis, P. Indyk, and R. Motwani, “Maintaining stream
statistics over sliding windows,” SIAM J. Computing 31, pp. 1794–1813,
2002.
5. P. Flajolet and G.N. Martin, “Probabilistic counting for database applica-
tions,” 24th Symposium on Foundations of Computer Science, pp. 76–82,
1983.
6. M. Garofalakis, J. Gehrke, and R. Rastogi (editors), Data Stream Man-
agement, Springer, 2009.
7. P.B. Gibbons, “Distinct sampling for highly-accurate answers to distinct
values queries and event reports,” Intl. Conf. on Very Large Databases,
pp. 541–550, 2001.
8. H.V. Jagadish, I.S. Mumick, and A. Silberschatz, “View maintenance
issues for the chronicle data model,” Proc. ACM Symp. on Principles of
Database Systems, pp. 113–124, 1995.
9. W.H. Kautz and R.C. Singleton, “Nonadaptive binary superimposed codes,”
IEEE Transactions on Information Theory 10, pp. 363–377, 1964.
10. J. Vitter, “Random sampling with a reservoir,” ACM Transactions on
Mathematical Software 11:1, pp. 37–57, 1985.
M
a
r
c
h
2
6
,
2
0
2
4
/
D
r
.
R
S](https://image.slidesharecdn.com/kit-601lecturenotes-unit-3-240523065037-f5d66706/75/KIT-601-Lecture-Notes-UNIT-3-pdf-Mining-Data-Stream-48-2048.jpg)
![M
a
r
c
h
2
6
,
2
0
2
4
/
D
r
.
R
S
11 Reference
[1] https://www.jigsawacademy.com/blogs/hr-analytics/data-analytics-lifecycle/
[2] https://statacumen.com/teach/ADA1/ADA1_notes_F14.pdf
[3] https://www.youtube.com/watch?v=fDRa82lxzaU
[4] https://www.investopedia.com/terms/d/data-analytics.asp
[5] http://egyankosh.ac.in/bitstream/123456789/10935/1/Unit-2.pdf
[6] http://epgp.inflibnet.ac.in/epgpdata/uploads/epgp_content/computer_science/16._d
ata_analytics/03._evolution_of_analytical_scalability/et/9280_et_3_et.pdf
[7] https://bhavanakhivsara.files.wordpress.com/2018/06/data-science-and-big-data-analy
-nieizv_book.pdf
[8] https://www.researchgate.net/publication/317214679_Sentiment_Analysis_for_Effect
ive_Stock_Market_Prediction
[9] https://snscourseware.org/snscenew/files/1569681518.pdf
[10] http://csis.pace.edu/ctappert/cs816-19fall/books/2015DataScienceBigDataAnalytics.
pdf
[11] https://www.youtube.com/watch?v=mccsmoh2_3c
[12] https://mentalmodels4life.net/2015/11/18/agile-data-science-applying-kanban-in-the-a
nalytics-life-cycle/
[13] https://www.sas.com/en_in/insights/big-data/what-is-big-data.html#:~:text=Big%20dat
a%20refers%20to%20data,around%20for%20a%20long%20time.
[14] https://www.javatpoint.com/big-data-characteristics
[15] Liu, S., Wang, M., Zhan, Y., Shi, J. (2009). Daily work stress and alcohol use: Testing the cross-
level moderation effects of neuroticism and job involvement. Personnel Psychology,62(3), 575–597.
http://dx.doi.org/10.1111/j.1744-6570.2009.01149.x
53](https://image.slidesharecdn.com/kit-601lecturenotes-unit-3-240523065037-f5d66706/75/KIT-601-Lecture-Notes-UNIT-3-pdf-Mining-Data-Stream-53-2048.jpg)
![M
a
r
c
h
2
6
,
2
0
2
4
/
D
r
.
R
S
[16] https://www.google.com/search?q=architecture+of+data+stream+modelsxsrf=APwXEdf9LJ8N
XMypRU-Sg28SH8m_pwiUDA:1679823244352source=lnmstbm=ischsa=Xved=2ahUKEwjGgY-epfn9
AhX5xTgGHRWjDmMQ_AUoAXoECAEQAwbiw=1366bih=622#imgrc=wnFWJQ01p-w_jM
********************
54](https://image.slidesharecdn.com/kit-601lecturenotes-unit-3-240523065037-f5d66706/75/KIT-601-Lecture-Notes-UNIT-3-pdf-Mining-Data-Stream-54-2048.jpg)
![M
a
r
c
h
2
6
,
2
0
2
4
/
D
r
.
R
S
processing system.
– Stream Processing Engine:— This is the core component of the stream processing system,
which processes the incoming stream data. The processing engine is responsible for ingesting
data2
, applying transformations, aggregating data, and producing output data in real-time.
Output Sinks:— These are the destinations of the processed stream data. The output sinks
may include databases, message queues, dashboards, or other systems that consume the output
data.
Figure 1: Illustration of Architecture of Data Stream Model [16].
The architecture (See Figure 1) of a stream processing system can be implemented using a variety of
technologies, such as message queues, stream processing frameworks, or microservices.
ˆ Stream Processing Frameworks—Stream processing frameworks are a popular choice for imple-
menting stream processing systems. These frameworks provide a set of libraries, APIs, and tools for
2Data ingestion is the process of importing large, assorted data files from multiple sources into a single, cloud-based storage
medium—a data warehouse, data mart or database—where it can be accessed and analyzed.
5](https://crownmelresort.com/image.slidesharecdn.com/kit-601lecturenotes-unit-3-240523065037-f5d66706/75/KIT-601-Lecture-Notes-UNIT-3-pdf-Mining-Data-Stream-5-2048.jpg)
![4.9. REFERENCES FOR CHAPTER 4 161
✦ Estimating the Number of 1’s in a Window: We can estimate the number
of 1’s in a window of 0’s and 1’s by grouping the 1’s into buckets. Each
bucket has a number of 1’s that is a power of 2; there are one or two
buckets of each size, and sizes never decrease as we go back in time. If
we record only the position and size of the buckets, we can represent the
contents of a window of size N with O(log2
N) space.
✦ Answering Queries About Numbers of 1’s: If we want to know the approx-
imate numbers of 1’s in the most recent k elements of a binary stream,
we find the earliest bucket B that is at least partially within the last k
positions of the window and estimate the number of 1’s to be the sum of
the sizes of each of the more recent buckets plus half the size of B. This
estimate can never be off by more that 50% of the true count of 1’s.
✦ Closer Approximations to the Number of 1’s: By changing the rule for
how many buckets of a given size can exist in the representation of a
binary window, so that either r or r − 1 of a given size may exist, we can
assure that the approximation to the true number of 1’s is never off by
more than 1/r.
✦ Exponentially Decaying Windows: Rather than fixing a window size, we
can imagine that the window consists of all the elements that ever arrived
in the stream, but with the element that arrived t time units ago weighted
by e−ct
for some time-constant c. Doing so allows us to maintain certain
summaries of an exponentially decaying window easily. For instance, the
weighted sum of elements can be recomputed, when a new element arrives,
by multiplying the old sum by 1 − c and then adding the new element.
✦ Maintaining Frequent Elements in an Exponentially Decaying Window:
We can imagine that each item is represented by a binary stream, where
0 means the item was not the element arriving at a given time, and 1
means that it was. We can find the elements whose sum of their binary
stream is at least 1/2. When a new element arrives, multiply all recorded
sums by 1 minus the time constant, add 1 to the count of the item that
just arrived, and delete from the record any item whose sum has fallen
below 1/2.
4.9 References for Chapter 4
Many ideas associated with stream management appear in the “chronicle data
model” of [8]. An early survey of research in stream-management systems is
[2]. Also, [6] is a recent book on the subject of stream management.
The sampling technique of Section 4.2 is from [7]. The Bloom Filter is
generally attributed to [3], although essentially the same technique appeared as
“superimposed codes” in [9].
M
a
r
c
h
2
6
,
2
0
2
4
/
D
r
.
R
S](https://crownmelresort.com/image.slidesharecdn.com/kit-601lecturenotes-unit-3-240523065037-f5d66706/75/KIT-601-Lecture-Notes-UNIT-3-pdf-Mining-Data-Stream-47-2048.jpg)
![162 CHAPTER 4. MINING DATA STREAMS
The algorithm for counting distinct elements is essentially that of [5], al-
though the particular method we described appears in [1]. The latter is also
the source for the algorithm for calculating the surprise number and higher
moments. However, the technique for maintaining a uniformly chosen sample
of positions in the stream is called “reservoir sampling” and comes from [10].
The technique for approximately counting 1’s in a window is from [4].
1. N. Alon, Y. Matias, and M. Szegedy, “The space complexity of approxi-
mating frequency moments,” 28th ACM Symposium on Theory of Com-
puting, pp. 20–29, 1996.
2. B. Babcock, S. Babu, M. Datar, R. Motwani, and J. Widom, “Models
and issues in data stream systems,” Symposium on Principles of Database
Systems, pp. 1–16, 2002.
3. B.H. Bloom, “Space/time trade-offs in hash coding with allowable errors,”
Comm. ACM 13:7, pp. 422–426, 1970.
4. M. Datar, A. Gionis, P. Indyk, and R. Motwani, “Maintaining stream
statistics over sliding windows,” SIAM J. Computing 31, pp. 1794–1813,
2002.
5. P. Flajolet and G.N. Martin, “Probabilistic counting for database applica-
tions,” 24th Symposium on Foundations of Computer Science, pp. 76–82,
1983.
6. M. Garofalakis, J. Gehrke, and R. Rastogi (editors), Data Stream Man-
agement, Springer, 2009.
7. P.B. Gibbons, “Distinct sampling for highly-accurate answers to distinct
values queries and event reports,” Intl. Conf. on Very Large Databases,
pp. 541–550, 2001.
8. H.V. Jagadish, I.S. Mumick, and A. Silberschatz, “View maintenance
issues for the chronicle data model,” Proc. ACM Symp. on Principles of
Database Systems, pp. 113–124, 1995.
9. W.H. Kautz and R.C. Singleton, “Nonadaptive binary superimposed codes,”
IEEE Transactions on Information Theory 10, pp. 363–377, 1964.
10. J. Vitter, “Random sampling with a reservoir,” ACM Transactions on
Mathematical Software 11:1, pp. 37–57, 1985.
M
a
r
c
h
2
6
,
2
0
2
4
/
D
r
.
R
S](https://crownmelresort.com/image.slidesharecdn.com/kit-601lecturenotes-unit-3-240523065037-f5d66706/75/KIT-601-Lecture-Notes-UNIT-3-pdf-Mining-Data-Stream-48-2048.jpg)
![M
a
r
c
h
2
6
,
2
0
2
4
/
D
r
.
R
S
11 Reference
[1] https://www.jigsawacademy.com/blogs/hr-analytics/data-analytics-lifecycle/
[2] https://statacumen.com/teach/ADA1/ADA1_notes_F14.pdf
[3] https://www.youtube.com/watch?v=fDRa82lxzaU
[4] https://www.investopedia.com/terms/d/data-analytics.asp
[5] http://egyankosh.ac.in/bitstream/123456789/10935/1/Unit-2.pdf
[6] http://epgp.inflibnet.ac.in/epgpdata/uploads/epgp_content/computer_science/16._d
ata_analytics/03._evolution_of_analytical_scalability/et/9280_et_3_et.pdf
[7] https://bhavanakhivsara.files.wordpress.com/2018/06/data-science-and-big-data-analy
-nieizv_book.pdf
[8] https://www.researchgate.net/publication/317214679_Sentiment_Analysis_for_Effect
ive_Stock_Market_Prediction
[9] https://snscourseware.org/snscenew/files/1569681518.pdf
[10] http://csis.pace.edu/ctappert/cs816-19fall/books/2015DataScienceBigDataAnalytics.
pdf
[11] https://www.youtube.com/watch?v=mccsmoh2_3c
[12] https://mentalmodels4life.net/2015/11/18/agile-data-science-applying-kanban-in-the-a
nalytics-life-cycle/
[13] https://www.sas.com/en_in/insights/big-data/what-is-big-data.html#:~:text=Big%20dat
a%20refers%20to%20data,around%20for%20a%20long%20time.
[14] https://www.javatpoint.com/big-data-characteristics
[15] Liu, S., Wang, M., Zhan, Y., Shi, J. (2009). Daily work stress and alcohol use: Testing the cross-
level moderation effects of neuroticism and job involvement. Personnel Psychology,62(3), 575–597.
http://dx.doi.org/10.1111/j.1744-6570.2009.01149.x
53](https://crownmelresort.com/image.slidesharecdn.com/kit-601lecturenotes-unit-3-240523065037-f5d66706/75/KIT-601-Lecture-Notes-UNIT-3-pdf-Mining-Data-Stream-53-2048.jpg)
![M
a
r
c
h
2
6
,
2
0
2
4
/
D
r
.
R
S
[16] https://www.google.com/search?q=architecture+of+data+stream+modelsxsrf=APwXEdf9LJ8N
XMypRU-Sg28SH8m_pwiUDA:1679823244352source=lnmstbm=ischsa=Xved=2ahUKEwjGgY-epfn9
AhX5xTgGHRWjDmMQ_AUoAXoECAEQAwbiw=1366bih=622#imgrc=wnFWJQ01p-w_jM
********************
54](https://crownmelresort.com/image.slidesharecdn.com/kit-601lecturenotes-unit-3-240523065037-f5d66706/75/KIT-601-Lecture-Notes-UNIT-3-pdf-Mining-Data-Stream-54-2048.jpg)
More Related Content
Similar to KIT-601 Lecture Notes-UNIT-3.pdf Mining Data Stream
![[WSO2Con EU 2018] Patterns for Building Streaming Apps](https://cdn.slidesharecdn.com/ss_thumbnails/3-181113092919-thumbnail.jpg?width=640&height=640&fit=bounds)
KIT-601 Lecture Notes-UNIT-3.pdf Mining Data Stream
- 1. M a r c h 2 6 , 2 0 2 4 / D r . R S Data Analytics (KIT-601) Unit-3:Mining Data Streams Dr. Radhey Shyam Professor Department of Information Technology SRMCEM Lucknow (Affiliated to Dr. A.P.J. Abdul Kalam Technical University, Lucknow) Unit-3 has been prepared and compiled by Dr. Radhey Shyam, with grateful acknowledgment to those who made their course contents freely available or (Contributed directly or indirectly). Feel free to use this study material for your own academic purposes. For any query, communication can be made through this email : shyam0058@gmail.com. March 26, 2024
- 2. Data Analytics (KIT601) Course Outcome ( CO) Bloom’s Knowledge Level (KL) At the end of course , the student will be able to CO 1 Discuss various concepts of data analytics pipeline K1, K2 CO 2 Apply classification and regression techniques K3 CO 3 Explain and apply mining techniques on streaming data K2, K3 CO 4 Compare different clustering and frequent pattern mining algorithms K4 CO 5 Describe the concept of R programming and implement analytics on Big data using R. K2,K3 DETAILED SYLLABUS 3-0-0 Unit Topic Proposed Lecture I Introduction to Data Analytics: Sources and nature of data, classification of data (structured, semi-structured, unstructured), characteristics of data, introduction to Big Data platform, need of data analytics, evolution of analytic scalability, analytic process and tools, analysis vs reporting, modern data analytic tools, applications of data analytics. Data Analytics Lifecycle: Need, key roles for successful analytic projects, various phases of data analytics lifecycle – discovery, data preparation, model planning, model building, communicating results, operationalization. 08 II Data Analysis: Regression modeling, multivariate analysis, Bayesian modeling, inference and Bayesian networks, support vector and kernel methods, analysis of time series: linear systems analysis & nonlinear dynamics, rule induction, neural networks: learning and generalisation, competitive learning, principal component analysis and neural networks, fuzzy logic: extracting fuzzy models from data, fuzzy decision trees, stochastic search methods. 08 III Mining Data Streams: Introduction to streams concepts, stream data model and architecture, stream computing, sampling data in a stream, filtering streams, counting distinct elements in a stream, estimating moments, counting oneness in a window, decaying window, Real-time Analytics Platform ( RTAP) applications, Case studies – real time sentiment analysis, stock market predictions. 08 IV Frequent Itemsets and Clustering: Mining frequent itemsets, market based modelling, Apriori algorithm, handling large data sets in main memory, limited pass algorithm, counting frequent itemsets in a stream, clustering techniques: hierarchical, K-means, clustering high dimensional data, CLIQUE and ProCLUS, frequent pattern based clustering methods, clustering in non-euclidean space, clustering for streams and parallelism. 08 V Frame Works and Visualization: MapReduce, Hadoop, Pig, Hive, HBase, MapR, Sharding, NoSQL Databases, S3, Hadoop Distributed File Systems, Visualization: visual data analysis techniques, interaction techniques, systems and applications. Introduction to R - R graphical user interfaces, data import and export, attribute and data types, descriptive statistics, exploratory data analysis, visualization before analysis, analytics for unstructured data. 08 Text books and References: 1. Michael Berthold, David J. Hand, Intelligent Data Analysis, Springer 2. Anand Rajaraman and Jeffrey David Ullman, Mining of Massive Datasets, Cambridge University Press. 3. John Garrett,Data Analytics for IT Networks : Developing Innovative Use Cases, Pearson Education Curriculum & Evaluation Scheme IT & CSI (V & VI semester) 23 M a r c h 2 6 , 2 0 2 4 / D r . R S
- 3. M a r c h 2 6 , 2 0 2 4 / D r . R S Unit-III: Mining DataStreams 1 Introduction to streams concepts Streams are a powerful concept in software development used for processing large amounts of data efficiently. A stream is a sequence of data that flows through a program from a source, such as a file or network connection. Streams are typically used to avoid loading the entire dataset into memory, which can be inefficient and slow down processing. In programming, there are two types of streams: input streams and output streams. Input streams are used to read data from a source, while output streams are used to write data to a destination. Both types of streams are typically used together to process data. Stream operations are used to manipulate data in a stream. The most common stream operations are filtering, mapping, and reducing. Filtering operations remove elements from a stream based on a condi- tion, mapping operations transform each element in a stream into a new element, and reducing operations aggregate elements in a stream into a single result. Streams are evaluated lazily, which means that data is only processed when it is needed. This allows programs to work with large datasets efficiently, as only the necessary data is processed at any given time. A stream pipeline is a series of stream operations that are applied to a stream. The output of one operation is used as the input for the next operation. Stream pipelines are used to transform data in a stream into a final result. Parallel streams are a type of stream that can be processed in parallel, using multiple threads. This can improve performance on multicore processors or in distributed computing environments. Java streams are a feature introduced in Java 8 that provide a concise and efficient way to work with collections and other data sources. Java streams include a variety of methods for filtering, mapping, and reducing data, and can be used to process data in parallel. Overall, streams are a powerful and flexible way to process data efficiently and effectively, making them an essential concept for software developers to understand. 3
- 4. M a r c h 2 6 , 2 0 2 4 / D r . R S 1.1 Stream DataModel and Architecture Stream data model and architecture1 are important concepts in stream processing. A stream data model defines the structure of the data being processed, while stream architecture refers to the system architecture used to process the stream data. ˆ Stream Data Model—The stream data model defines the structure of the data being processed in a stream. Typically, stream data is represented as a series of records, where each record contains one or more fields. The fields in a record represent the attributes of the data being processed. The stream data model is typically implemented using a schema, which defines the structure and data types of the fields in each record. The schema is used to validate incoming data and ensure that it conforms to the expected structure. ˆ Stream Architecture—The architecture of a stream processing system defines the components and infrastructure used to process stream data. Stream processing systems are typically composed of three main components: – Data Sources:— These are the sources of the stream data, such as sensors, log files, or message queues. The data sources generate data in real-time, which is then processed by the stream 1The architecture of a data stream model consists of three main components: sources, processing units, and sinks. Sources: Sources are the origin points of data streams. They can be sensors, IoT devices, social media platforms, or any other source that generates a continuous stream of data. Sources can send data to processing units either directly or through message brokers. Processing Units: Processing units are the main components of the data stream model. They receive data from sources, process it, and send it to the next processing unit or sink. Processing units can be classified into three types: streaming engines, data processing engines, and storage engines. Streaming engines: Streaming engines are responsible for ingesting data from sources and delivering it to the next processing unit or sink. Examples of streaming engines are Apache Kafka, Apache Flink, and Apache Storm. Data processing engines: Data processing engines are responsible for processing the data and applying various operations such as filtering, aggregation, and transformation. Examples of data processing engines are Apache Spark and Apache Beam. Storage engines: Storage engines are responsible for storing the data stream or the results of data processing operations. Examples of storage engines are Apache Cassandra and Apache HBase. Sinks: Sinks are the final destinations of data streams. They can be databases, visualization tools, or any other system that consumes data. Sinks receive data from processing units and store or display it. In addition to these three main components, the data stream model may also include message brokers, which act as intermediaries between sources and processing units, and connectors, which facilitate the integration of different systems in the data stream model. Overall, the architecture (See Figure 1) of a data stream model is designed to handle high-volume, high-velocity, and high- variety data streams in real-time, making it ideal for applications such as real-time analytics, online monitoring, and IoT data processing. 4
- 5. M a r c h 2 6 , 2 0 2 4 / D r . R S processing system. – StreamProcessing Engine:— This is the core component of the stream processing system, which processes the incoming stream data. The processing engine is responsible for ingesting data2 , applying transformations, aggregating data, and producing output data in real-time. Output Sinks:— These are the destinations of the processed stream data. The output sinks may include databases, message queues, dashboards, or other systems that consume the output data. Figure 1: Illustration of Architecture of Data Stream Model [16]. The architecture (See Figure 1) of a stream processing system can be implemented using a variety of technologies, such as message queues, stream processing frameworks, or microservices. ˆ Stream Processing Frameworks—Stream processing frameworks are a popular choice for imple- menting stream processing systems. These frameworks provide a set of libraries, APIs, and tools for 2Data ingestion is the process of importing large, assorted data files from multiple sources into a single, cloud-based storage medium—a data warehouse, data mart or database—where it can be accessed and analyzed. 5
- 6. M a r c h 2 6 , 2 0 2 4 / D r . R S building, deploying, andmanaging stream processing applications. Some popular stream processing frameworks include Apache Kafka, Apache Flink, and Apache Spark. Stream processing frameworks typically provide the following features: – Data ingestion: Stream processing frameworks provide APIs and connectors for ingesting data from various sources, such as message queues, databases, or file systems. – Data transformation: Stream processing frameworks provide APIs and operators for trans- forming the incoming data, such as filtering, mapping, or aggregating data. – Real-time processing: Stream processing frameworks are designed for real-time processing of stream data, allowing for low-latency processing and real-time analysis of data. – Scalability: Stream processing frameworks are designed to scale horizontally, allowing for pro- cessing of large volumes of data and high-throughput processing. Overall, stream data model and architecture are important concepts for stream processing systems, providing a flexible and efficient way to process large volumes of data in real-time. 2 Stream Computing Stream computing is a type of data processing that involves the continuous and real-time processing of data streams. Data streams can come from various sources, such as social media feeds, sensor data, or financial market data, and are typically unbounded and constantly changing. Stream computing systems are designed to process these data streams quickly and efficiently, often in parallel, to provide real-time insights and decision-making capabilities. These systems typically use distributed computing technologies and are capable of handling large volumes of data in real time. Stream computing is used in a variety of applications, including fraud detection, real-time analytics, online recommendations, and monitoring of Internet of Things (IoT) devices. Some popular stream computing platforms include Apache Kafka, Apache Storm, and Apache Flink. Detailed explanation of stream computing: 1. Real-time processing: Stream computing is all about processing data in real-time. Data streams are continuously flowing in and need to be processed as they arrive. This requires processing engines that can handle data as it comes in, rather than storing and processing it later. 6
- 7. M a r c h 2 6 , 2 0 2 4 / D r . R S 2. Unbounded datastreams: Stream computing involves processing unbounded data streams, meaning that the data never stops flowing. This requires systems that can handle large volumes of data and scale up or down as needed. 3. Distributed computing: Stream computing systems typically use distributed computing technolo- gies, where the processing workload is distributed across multiple machines. This allows for parallel processing of data streams, which can improve performance and scalability. 4. Fault tolerance: Stream computing systems need to be fault-tolerant, meaning that they can continue to function even if some components fail. This is especially important for systems that are processing critical data or providing real-time insights that are needed for decision-making. 5. Data processing algorithms: Stream computing involves using algorithms that can process data streams efficiently and accurately. These algorithms need to be designed to handle unbounded data streams, and to provide real-time insights and decision-making capabilities. 6. Applications: Stream computing has a wide range of applications, including fraud detection, real- time analytics, online recommendations, and monitoring of IoT devices. In each of these cases, stream computing can provide real-time insights that can help businesses make better decisions or react quickly to changing circumstances. 7. Platforms: There are many different stream computing platforms available, including open source solutions like Apache Kafka, Apache Storm, and Apache Flink, as well as commercial solutions from companies like IBM and Microsoft. Overall, stream computing is a powerful approach to processing data in real-time, allowing businesses to make better decisions and respond quickly to changing circumstances. 3 Sampling Data in a Stream Sampling data in a stream involves selecting a subset of the data that represents the overall characteristics of the entire stream. This can be useful in cases where processing the entire data stream in real-time is not feasible, either due to processing limitations or resource constraints. Here are some common techniques for sampling data in a stream: 7
- 8. M a r c h 2 6 , 2 0 2 4 / D r . R S 1. Random Sampling:This technique involves randomly selecting a subset of data points from the data stream. This can be done using a random number generator or a hash function that maps each data point to a random number between 0 and 1. The probability of selecting a given data point is equal to its weight in the hash function. Random sampling can be useful when the data stream is evenly distributed, and each data point has an equal probability of being selected. 2. Systematic Sampling: This technique involves selecting every nth data point from the stream, where n is a predetermined constant. Systematic sampling can be useful when the data stream is periodic, and the sampling interval aligns with the period of the data stream. 3. Stratified Sampling: This technique involves dividing the data stream into strata or subgroups based on certain characteristics, such as data type, source, or time period. A random or systematic sampling method can then be applied to each stratum to select a representative sample. Stratified sampling can be useful when the data stream is heterogeneous, and the characteristics of each subgroup are known. 4. Reservoir Sampling: This technique involves maintaining a reservoir or buffer of a fixed size that stores a random subset of the data points from the stream. As new data points arrive, they are compared to the data points in the reservoir, and if they have a higher probability of being selected, they are added to the reservoir, and a random data point is removed. Reservoir sampling can be useful when the data stream is of unknown size, and the sample needs to be representative of the entire data stream. Sampling data in a stream requires careful consideration of the sampling technique used, as well as the size and representativeness of the sample. Additionally, it’s important to consider the trade-offs between processing speed and accuracy when selecting a sampling technique. 4 Filtering Streams Filtering streams involves selecting a subset of the data in a stream that meets certain criteria or conditions. This can be useful in cases where the entire data stream is not relevant or useful, and only a portion of the data needs to be processed or analyzed. Here are some common techniques for filtering streams: 8
- 9. M a r c h 2 6 , 2 0 2 4 / D r . R S 1. Time-based filtering:This technique involves selecting data points from a specific time period or window. For example, only data points that fall within a specific hour of the day or day of the week may be selected. Time-based filtering can be useful when analyzing trends or patterns over time. 2. Attribute-based filtering: This technique involves selecting data points based on certain attributes or characteristics. For example, only data points that meet a certain threshold or range for a particular attribute, such as temperature or volume, may be selected. Attribute-based filtering can be useful when looking for specific patterns or anomalies in the data. 3. Pattern-based filtering: This technique involves selecting data points that match a specific pattern or sequence. For example, only data points that follow a certain sequence of events or meet a specific condition may be selected. Pattern-based filtering can be useful when analyzing data from sensors or other IoT devices. 4. Keyword-based filtering: This technique involves selecting data points that contain specific key- words or phrases. For example, only social media posts that contain a certain hashtag or keyword may be selected. Keyword-based filtering can be useful for sentiment analysis or tracking social media trends. Filtering streams can be a powerful technique for processing and analyzing data in real-time. It can help reduce the amount of data that needs to be processed, making it easier to handle large volumes of data and reduce processing time. However, it’s important to carefully consider the filtering criteria used and ensure that the resulting data subset is still representative of the entire data stream. 5 Counting Distinct Elements in a Stream Counting distinct elements in a stream involves determining the number of unique elements that appear in a data stream. This can be useful in cases where duplicate or repeated elements are not relevant or useful, and only the unique elements need to be analyzed. Here are some common techniques for counting distinct elements in a stream: 1. Hash-based techniques: This technique involves using a hash function to map each data point in the stream to a fixed-size hash code. The hash codes can then be stored in a hash table or data structure, and duplicate hash codes can be ignored. Hash-based techniques can be useful for processing large data streams with a low number of distinct elements. 9
- 10. M a r c h 2 6 , 2 0 2 4 / D r . R S 2. Count-min sketch:This technique involves using a probabilistic data structure called a Count-min sketch to estimate the frequency of each data point in the stream. The Count-min sketch uses a hash function to map each data point to a series of counters, and the minimum counter value is used to estimate the frequency of each data point. The Count-min sketch can be useful for processing large data streams with a high number of distinct elements. 3. Bloom filter: This technique involves using a probabilistic data structure called a Bloom filter to test whether a data point is likely to be in the set of distinct elements. The Bloom filter uses a hash function to map each data point to a series of bits, and the bits are set to 1 in the Bloom filter if the corresponding hash code is generated. The Bloom filter can be useful for processing large data streams with a high number of distinct elements when memory usage is a concern. Counting distinct elements in a stream requires careful consideration of the technique used, as well as the size and complexity of the data stream. Additionally, it’s important to consider the trade-offs between accuracy and memory usage when selecting a technique, as some techniques may sacrifice accuracy for memory efficiency. 6 Estimating Moments Estimating moments in a data stream involves determining statistical properties of the data distribution based on a subset of the data. Moments are a set of summary statistics that describe the distribution of a dataset, including measures such as the mean, variance, skewness3 , and kurtosis4 . Here are some common techniques for estimating moments in a data stream: 1. Sample mean and variance: This technique involves computing the sample mean and variance of a subset of the data stream to estimate the population mean and variance. The sample mean is the average of the subset of data points, while the sample variance measures the spread of the data points around the sample mean. This technique is simple and efficient, but can be sensitive to outliers and changes in the data distribution. Count-min sketch: This technique can also be used to estimate moments in a data stream, such as the frequency of a data point or the number of distinct elements in the stream. The Count-min sketch 3Skewness, in statistics, is the degree of asymmetry observed in a probability distribution. 4Kurtosis is a measure of the tailedness of a distribution. Tailedness is how often outliers occur. 10
- 11. M a r c h 2 6 , 2 0 2 4 / D r . R S uses a hashfunction to map each data point to a series of counters, and the minimum counter value is used to estimate the frequency of each data point. The Count-min sketch can be useful for processing large data streams with a high number of distinct elements. Streaming algorithms: There are several streaming algorithms that can be used to estimate mo- ments in a data stream, including the AMS algorithm and the Greenwald-Khanna algorithm. These algorithms use a sample of the data stream to estimate the moments of the full data stream. These algorithms are designed to handle large data streams with a low memory footprint. Estimating moments in a data stream can be useful for summarizing large datasets and providing insights into the underlying data distribution. However, it’s important to consider the trade-offs between accuracy and memory usage when selecting a technique, as some techniques may sacrifice accuracy for memory efficiency. 7 Counting Oneness in a Window Counting oneness in a window of a data stream involves counting the number of unique or distinct elements that appear within a fixed-size window of the stream. This can be useful for identifying patterns or trends in the data over time, or for detecting anomalies or outliers in the data. Here are some common techniques for counting oneness in a window of a data stream: 1. Sliding window technique: This technique involves dividing the data stream into fixed-size windows and counting the number of unique elements in each window. The window size can be adjusted based on the characteristics of the data stream and the desired level of granularity. This technique is simple and efficient, but can be sensitive to changes in the data distribution over time. 2. Bloom filter: This technique involves using a probabilistic data structure called a Bloom filter to test whether a data point is likely to be in the set of unique elements within a window. The Bloom filter uses a hash function to map each data point to a series of bits, and the bits are set to 1 in the Bloom filter if the corresponding hash code is generated. The Bloom filter can be useful for processing large data streams with a high number of distinct elements when memory usage is a concern. 3. Count-min sketch: This technique involves using a probabilistic data structure called a Count-min sketch to estimate the frequency of each data point in a window. The Count-min sketch uses a hash function to map each data point to a series of counters, and the minimum counter value is used to 11
- 12. M a r c h 2 6 , 2 0 2 4 / D r . R S estimate the frequencyof each data point. The Count-min sketch can be useful for processing large data streams with a high number of distinct elements. Counting oneness in a window of a data stream can be useful for understanding the unique patterns and trends in the data over time. However, it’s important to consider the trade-offs between accuracy and memory usage when selecting a technique, as some techniques may sacrifice accuracy for memory efficiency. Additionally, it’s important to carefully choose the size of the window to capture relevant patterns in the data while minimizing noise or irrelevant data. 8 Decaying Window A decaying window in a data stream is a technique for processing a stream of data by giving more weight to recent data points while gradually reducing the weight of older data points. This technique is useful for analyzing data streams that exhibit time-varying patterns or trends, where recent data points may be more relevant than older ones. Here are some common techniques for implementing a decaying window in a data stream: 1. Exponential decay: This technique involves assigning exponentially decreasing weights to data points based on their age in the window. The weight of a data point is given by a function of the form wi = alpha ∗ (1 − beta)i , where i is the index of the data point in the window, alpha is a scaling factor, and beta is a decay rate parameter. This technique gives more weight to recent data points while gradually reducing the weight of older ones. 2. Moving average: This technique involves computing the average of a sliding window of data points, where the weights of the data points are given by a weighted function that gives more weight to recent data points. The weights of the data points can be determined based on their age in the window, or by using an exponential decay function as described above. 3. Forgetfulness mechanism: This technique involves gradually removing older data points from the window to give more weight to recent data points. This can be achieved by using a data structure that supports efficient deletion of old data points, such as a priority queue or a heap. Decaying windows can be useful for analyzing time-varying data streams, such as sensor data or stock prices, where recent data points may be more relevant than older ones. However, it’s important to carefully 12
- 13. M a r c h 2 6 , 2 0 2 4 / D r . R S choose the decayrate and window size to balance the need for responsiveness to recent data with the desire to capture long-term trends and patterns in the data. Additionally, it’s important to consider the computational complexity of the chosen technique and its impact on memory usage and processing time. 9 Real-time Analytics Platform (RTAP) Applications Real-time Analytics Platform (RTAP) applications are software applications that allow businesses to analyze and visualize data in real-time. These applications are designed to process large volumes of data in real-time, providing businesses with immediate insights into their operations, customer behavior, and market trends. Here are some examples of RTAP applications: 1. E-commerce: RTAP applications can be used to analyze customer behavior and preferences in real- time, providing businesses with insights into customer preferences and buying patterns. This can help businesses to personalize their offerings and improve customer satisfaction. 2. Financial services: RTAP applications can be used to analyze market trends and financial data in real-time, allowing businesses to make informed investment decisions and manage risk more effectively. 3. Social media: RTAP applications can be used to analyze social media data in real-time, providing businesses with insights into customer sentiment and preferences. This can help businesses to improve their marketing and advertising campaigns and enhance their brand reputation. 4. Healthcare: RTAP applications can be used to analyze patient data in real-time, providing healthcare providers with insights into patient health and treatment outcomes. This can help providers to improve patient care and reduce costs. 5. Transportation: RTAP applications can be used to analyze transportation data in real-time, pro- viding businesses with insights into traffic patterns, route optimization, and driver behavior. This can help businesses to improve their logistics and reduce transportation costs. Overall, RTAP applications can provide businesses with a competitive advantage by allowing them to make informed decisions in real-time. By processing and analyzing large volumes of data in real-time, businesses can gain valuable insights into their operations, customer behavior, and market trends, allowing them to make informed decisions that drive business success. 13
- 14. M a r c h 2 6 , 2 0 2 4 / D r . R S 10 Case studies 10.1Real Time Sentiment Analysis Real-time sentiment analysis is a powerful application of real-time analytics that allows businesses to monitor and analyze customer sentiment in real-time. Here are some case studies of companies that have successfully implemented real-time sentiment analysis to improve their customer engagement and satisfaction: ˆ American Airlines: American Airlines implemented a real-time sentiment analysis solution to moni- tor customer sentiment on social media platforms such as Twitter and Facebook. By analyzing customer feedback in real-time, the airline was able to identify customer issues and respond to them in a timely manner, improving customer satisfaction and reducing negative feedback. ˆ Domino’s Pizza: Domino’s Pizza implemented a real-time sentiment analysis solution to monitor customer feedback on social media platforms. By analyzing customer sentiment in real-time, the pizza chain was able to quickly identify and address customer issues, improving customer satisfaction and loyalty. ˆ Sprint: Sprint, a telecommunications company, implemented a real-time sentiment analysis solution to monitor customer feedback on social media platforms. By analyzing customer sentiment in real-time, Sprint was able to identify and address customer issues more quickly, improving customer satisfaction and reducing churn. ˆ Walmart: Walmart, a retail giant, implemented a real-time sentiment analysis solution to monitor customer feedback on social media platforms. By analyzing customer sentiment in real-time, Walmart was able to identify trends and patterns in customer feedback, allowing the company to improve its product offerings and customer service. ˆ Airbnb: Airbnb, an online marketplace for short-term rentals, implemented a real-time sentiment analysis solution to monitor customer feedback on social media platforms. By analyzing customer sentiment in real-time, Airbnb was able to identify areas for improvement in its customer service and product offerings, leading to higher customer satisfaction and loyalty. Overall, real-time sentiment analysis is a powerful tool for businesses to monitor and analyze customer sentiment in real-time. By analyzing customer feedback in real-time, businesses can identify issues and respond to them quickly, improving customer satisfaction and loyalty. 14
- 15. M a r c h 2 6 , 2 0 2 4 / D r . R S 10.2 Stock MarketPredictions Stock market predictions using real-time analytics is a complex and challenging application. However, here are some examples of companies that have successfully implemented real-time analytics to improve their stock market predictions: ˆ Bloomberg: Bloomberg is a financial information company that provides real-time data and analytics to financial professionals. The company uses real-time analytics to track market trends, news, and economic indicators, allowing traders and investors to make informed decisions in real-time. ˆ BlackRock: BlackRock, an investment management company, uses real-time analytics to analyze market data and predict stock market trends. The company uses machine learning algorithms to analyze vast amounts of data in real-time, allowing it to make more accurate predictions. ˆ Citadel: Citadel is a hedge fund that uses real-time analytics to analyze market data and make investment decisions in real-time. The company uses machine learning algorithms to analyze market trends and news, allowing it to make more informed investment decisions. ˆ Renaissance Technologies: Renaissance Technologies is a hedge fund that uses real-time analytics to analyze market data and predict stock market trends. The company uses machine learning algorithms to analyze vast amounts of data in real-time, allowing it to make more accurate predictions. ˆ JPMorgan: JPMorgan, a financial services company, uses real-time analytics to analyze market data and make investment decisions in real-time. The company uses machine learning algorithms to analyze market trends and news, allowing it to make more informed investment decisions. Overall, real-time analytics has revolutionized the way financial professionals analyze and predict stock mar- ket trends. By analyzing vast amounts of data in real-time, companies can make more informed investment decisions, leading to higher returns and greater financial success. 15
- 16.
- 17. Chapter 4 Mining DataStreams Most of the algorithms described in this book assume that we are mining a database. That is, all our data is available when and if we want it. In this chapter, we shall make another assumption: data arrives in a stream or streams, and if it is not processed immediately or stored, then it is lost forever. Moreover, we shall assume that the data arrives so rapidly that it is not feasible to store it all in active storage (i.e., in a conventional database), and then interact with it at the time of our choosing. The algorithms for processing streams each involve summarization of the stream in some way. We shall start by considering how to make a useful sample of a stream and how to filter a stream to eliminate most of the “undesirable” elements. We then show how to estimate the number of different elements in a stream using much less storage than would be required if we listed all the elements we have seen. Another approach to summarizing a stream is to look at only a fixed-length “window” consisting of the last n elements for some (typically large) n. We then query the window as if it were a relation in a database. If there are many streams and/or n is large, we may not be able to store the entire window for every stream, so we need to summarize even the windows. We address the fundamental problem of maintaining an approximate count on the number of 1’s in the window of a bit stream, while using much less space than would be needed to store the entire window itself. This technique generalizes to approximating various kinds of sums. 4.1 The Stream Data Model Let us begin by discussing the elements of streams and stream processing. We explain the difference between streams and databases and the special problems that arise when dealing with streams. Some typical applications where the stream model applies will be examined. 131 M a r c h 2 6 , 2 0 2 4 / D r . R S
- 18. 132 CHAPTER 4.MINING DATA STREAMS Queries Standing Stream Processor . . . 0, 1, 1, 0, 1, 0, 0, 0 q, w, e, r, t, y, u, i, o 1, 5, 2, 7, 4, 0, 3, 5 time Output streams Storage Storage Archival Streams entering Ad−hoc Queries Limited Working Figure 4.1: A data-stream-management system 4.1.1 A Data-Stream-Management System In analogy to a database-management system, we can view a stream processor as a kind of data-management system, the high-level organization of which is suggested in Fig. 4.1. Any number of streams can enter the system. Each stream can provide elements at its own schedule; they need not have the same data rates or data types, and the time between elements of one stream need not be uniform. The fact that the rate of arrival of stream elements is not under the control of the system distinguishes stream processing from the processing of data that goes on within a database-management system. The latter system controls the rate at which data is read from the disk, and therefore never has to worry about data getting lost as it attempts to execute queries. Streams may be archived in a large archival store, but we assume it is not possible to answer queries from the archival store. It could be examined only under special circumstances using time-consuming retrieval processes. There is also a working store, into which summaries or parts of streams may be placed, and which can be used for answering queries. The working store might be disk, or it might be main memory, depending on how fast we need to process queries. But either way, it is of sufficiently limited capacity that it cannot store all the data from all the streams. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 19. 4.1. THE STREAMDATA MODEL 133 4.1.2 Examples of Stream Sources Before proceeding, let us consider some of the ways in which stream data arises naturally. Sensor Data Imagine a temperature sensor bobbing about in the ocean, sending back to a base station a reading of the surface temperature each hour. The data produced by this sensor is a stream of real numbers. It is not a very interesting stream, since the data rate is so low. It would not stress modern technology, and the entire stream could be kept in main memory, essentially forever. Now, give the sensor a GPS unit, and let it report surface height instead of temperature. The surface height varies quite rapidly compared with tempera- ture, so we might have the sensor send back a reading every tenth of a second. If it sends a 4-byte real number each time, then it produces 3.5 megabytes per day. It will still take some time to fill up main memory, let alone a single disk. But one sensor might not be that interesting. To learn something about ocean behavior, we might want to deploy a million sensors, each sending back a stream, at the rate of ten per second. A million sensors isn’t very many; there would be one for every 150 square miles of ocean. Now we have 3.5 terabytes arriving every day, and we definitely need to think about what can be kept in working storage and what can only be archived. Image Data Satellites often send down to earth streams consisting of many terabytes of images per day. Surveillance cameras produce images with lower resolution than satellites, but there can be many of them, each producing a stream of images at intervals like one second. London is said to have six million such cameras, each producing a stream. Internet and Web Traffic A switching node in the middle of the Internet receives streams of IP packets from many inputs and routes them to its outputs. Normally, the job of the switch is to transmit data and not to retain it or query it. But there is a tendency to put more capability into the switch, e.g., the ability to detect denial-of-service attacks or the ability to reroute packets based on information about congestion in the network. Web sites receive streams of various types. For example, Google receives sev- eral hundred million search queries per day. Yahoo! accepts billions of “clicks” per day on its various sites. Many interesting things can be learned from these streams. For example, an increase in queries like “sore throat” enables us to track the spread of viruses. A sudden increase in the click rate for a link could M a r c h 2 6 , 2 0 2 4 / D r . R S
- 20. 134 CHAPTER 4.MINING DATA STREAMS indicate some news connected to that page, or it could mean that the link is broken and needs to be repaired. 4.1.3 Stream Queries There are two ways that queries get asked about streams. We show in Fig. 4.1 a place within the processor where standing queries are stored. These queries are, in a sense, permanently executing, and produce outputs at appropriate times. Example 4.1 : The stream produced by the ocean-surface-temperature sen- sor mentioned at the beginning of Section 4.1.2 might have a standing query to output an alert whenever the temperature exceeds 25 degrees centigrade. This query is easily answered, since it depends only on the most recent stream element. Alternatively, we might have a standing query that, each time a new reading arrives, produces the average of the 24 most recent readings. That query also can be answered easily, if we store the 24 most recent stream elements. When a new stream element arrives, we can drop from the working store the 25th most recent element, since it will never again be needed (unless there is some other standing query that requires it). Another query we might ask is the maximum temperature ever recorded by that sensor. We can answer this query by retaining a simple summary: the maximum of all stream elements ever seen. It is not necessary to record the entire stream. When a new stream element arrives, we compare it with the stored maximum, and set the maximum to whichever is larger. We can then answer the query by producing the current value of the maximum. Similarly, if we want the average temperature over all time, we have only to record two values: the number of readings ever sent in the stream and the sum of those readings. We can adjust these values easily each time a new reading arrives, and we can produce their quotient as the answer to the query. 2 The other form of query is ad-hoc, a question asked once about the current state of a stream or streams. If we do not store all streams in their entirety, as normally we can not, then we cannot expect to answer arbitrary queries about streams. If we have some idea what kind of queries will be asked through the ad-hoc query interface, then we can prepare for them by storing appropriate parts or summaries of streams as in Example 4.1. If we want the facility to ask a wide variety of ad-hoc queries, a common approach is to store a sliding window of each stream in the working store. A sliding window can be the most recent n elements of a stream, for some n, or it can be all the elements that arrived within the last t time units, e.g., one day. If we regard each stream element as a tuple, we can treat the window as a relation and query it with any SQL query. Of course the stream-management system must keep the window fresh, deleting the oldest elements as new ones come in. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 21. 4.1. THE STREAMDATA MODEL 135 Example 4.2 : Web sites often like to report the number of unique users over the past month. If we think of each login as a stream element, we can maintain a window that is all logins in the most recent month. We must associate the arrival time with each login, so we know when it no longer belongs to the window. If we think of the window as a relation Logins(name, time), then it is simple to get the number of unique users over the past month. The SQL query is: SELECT COUNT(DISTINCT(name)) FROM Logins WHERE time >= t; Here, t is a constant that represents the time one month before the current time. Note that we must be able to maintain the entire stream of logins for the past month in working storage. However, for even the largest sites, that data is not more than a few terabytes, and so surely can be stored on disk. 2 4.1.4 Issues in Stream Processing Before proceeding to discuss algorithms, let us consider the constraints under which we work when dealing with streams. First, streams often deliver elements very rapidly. We must process elements in real time, or we lose the opportunity to process them at all, without accessing the archival storage. Thus, it often is important that the stream-processing algorithm is executed in main memory, without access to secondary storage or with only rare accesses to secondary storage. Moreover, even when streams are “slow,” as in the sensor-data example of Section 4.1.2, there may be many such streams. Even if each stream by itself can be processed using a small amount of main memory, the requirements of all the streams together can easily exceed the amount of available main memory. Thus, many problems about streaming data would be easy to solve if we had enough memory, but become rather hard and require the invention of new techniques in order to execute them at a realistic rate on a machine of realistic size. Here are two generalizations about stream algorithms worth bearing in mind as you read through this chapter: • Often, it is much more efficient to get an approximate answer to our problem than an exact solution. • As in Chapter 3, a variety of techniques related to hashing turn out to be useful. Generally, these techniques introduce useful randomness into the algorithm’s behavior, in order to produce an approximate answer that is very close to the true result. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 22. 136 CHAPTER 4.MINING DATA STREAMS 4.2 Sampling Data in a Stream As our first example of managing streaming data, we shall look at extracting reliable samples from a stream. As with many stream algorithms, the “trick” involves using hashing in a somewhat unusual way. 4.2.1 A Motivating Example The general problem we shall address is selecting a subset of a stream so that we can ask queries about the selected subset and have the answers be statistically representative of the stream as a whole. If we know what queries are to be asked, then there are a number of methods that might work, but we are looking for a technique that will allow ad-hoc queries on the sample. We shall look at a particular problem, from which the general idea will emerge. Our running example is the following. A search engine receives a stream of queries, and it would like to study the behavior of typical users.1 We assume the stream consists of tuples (user, query, time). Suppose that we want to answer queries such as “What fraction of the typical user’s queries were repeated over the past month?” Assume also that we wish to store only 1/10th of the stream elements. The obvious approach would be to generate a random number, say an integer from 0 to 9, in response to each search query. Store the tuple if and only if the random number is 0. If we do so, each user has, on average, 1/10th of their queries stored. Statistical fluctuations will introduce some noise into the data, but if users issue many queries, the law of large numbers will assure us that most users will have a fraction quite close to 1/10th of their queries stored. However, this scheme gives us the wrong answer to the query asking for the average number of duplicate queries for a user. Suppose a user has issued s search queries one time in the past month, d search queries twice, and no search queries more than twice. If we have a 1/10th sample, of queries, we shall see in the sample for that user an expected s/10 of the search queries issued once. Of the d search queries issued twice, only d/100 will appear twice in the sample; that fraction is d times the probability that both occurrences of the query will be in the 1/10th sample. Of the queries that appear twice in the full stream, 18d/100 will appear exactly once. To see why, note that 18/100 is the probability that one of the two occurrences will be in the 1/10th of the stream that is selected, while the other is in the 9/10th that is not selected. The correct answer to the query about the fraction of repeated searches is d/(s+d). However, the answer we shall obtain from the sample is d/(10s+19d). To derive the latter formula, note that d/100 appear twice, while s/10+18d/100 appear once. Thus, the fraction appearing twice in the sample is d/100 divided 1While we shall refer to “users,” the search engine really receives IP addresses from which the search query was issued. We shall assume that these IP addresses identify unique users, which is approximately true, but not exactly true. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 23. 4.2. SAMPLING DATAIN A STREAM 137 by d/100 + s/10 + 18d/100. This ratio is d/(10s + 19d). For no positive values of s and d is d/(s + d) = d/(10s + 19d). 4.2.2 Obtaining a Representative Sample The query of Section 4.2.1, like many queries about the statistics of typical users, cannot be answered by taking a sample of each user’s search queries. Thus, we must strive to pick 1/10th of the users, and take all their searches for the sample, while taking none of the searches from other users. If we can store a list of all users, and whether or not they are in the sample, then we could do the following. Each time a search query arrives in the stream, we look up the user to see whether or not they are in the sample. If so, we add this search query to the sample, and if not, then not. However, if we have no record of ever having seen this user before, then we generate a random integer between 0 and 9. If the number is 0, we add this user to our list with value “in,” and if the number is other than 0, we add the user with the value “out.” That method works as long as we can afford to keep the list of all users and their in/out decision in main memory, because there isn’t time to go to disk for every search that arrives. By using a hash function, one can avoid keeping the list of users. That is, we hash each user name to one of ten buckets, 0 through 9. If the user hashes to bucket 0, then accept this search query for the sample, and if not, then not. Note we do not actually store the user in the bucket; in fact, there is no data in the buckets at all. Effectively, we use the hash function as a random- number generator, with the important property that, when applied to the same user several times, we always get the same “random” number. That is, without storing the in/out decision for any user, we can reconstruct that decision any time a search query by that user arrives. More generally, we can obtain a sample consisting of any rational fraction a/b of the users by hashing user names to b buckets, 0 through b − 1. Add the search query to the sample if the hash value is less than a. 4.2.3 The General Sampling Problem The running example is typical of the following general problem. Our stream consists of tuples with n components. A subset of the components are the key components, on which the selection of the sample will be based. In our running example, there are three components – user, query, and time – of which only user is in the key. However, we could also take a sample of queries by making query be the key, or even take a sample of user-query pairs by making both those components form the key. To take a sample of size a/b, we hash the key value for each tuple to b buckets, and accept the tuple for the sample if the hash value is less than a. If the key consists of more than one component, the hash function needs to combine the values for those components to make a single hash-value. The M a r c h 2 6 , 2 0 2 4 / D r . R S
- 24. 138 CHAPTER 4.MINING DATA STREAMS result will be a sample consisting of all tuples with certain key values. The selected key values will be approximately a/b of all the key values appearing in the stream. 4.2.4 Varying the Sample Size Often, the sample will grow as more of the stream enters the system. In our running example, we retain all the search queries of the selected 1/10th of the users, forever. As time goes on, more searches for the same users will be accumulated, and new users that are selected for the sample will appear in the stream. If we have a budget for how many tuples from the stream can be stored as the sample, then the fraction of key values must vary, lowering as time goes on. In order to assure that at all times, the sample consists of all tuples from a subset of the key values, we choose a hash function h from key values to a very large number of values 0, 1, . . . , B−1. We maintain a threshold t, which initially can be the largest bucket number, B − 1. At all times, the sample consists of those tuples whose key K satisfies h(K) ≤ t. New tuples from the stream are added to the sample if and only if they satisfy the same condition. If the number of stored tuples of the sample exceeds the allotted space, we lower t to t−1 and remove from the sample all those tuples whose key K hashes to t. For efficiency, we can lower t by more than 1, and remove the tuples with several of the highest hash values, whenever we need to throw some key values out of the sample. Further efficiency is obtained by maintaining an index on the hash value, so we can find all those tuples whose keys hash to a particular value quickly. 4.2.5 Exercises for Section 4.2 Exercise 4.2.1: Suppose we have a stream of tuples with the schema Grades(university, courseID, studentID, grade) Assume universities are unique, but a courseID is unique only within a uni- versity (i.e., different universities may have different courses with the same ID, e.g., “CS101”) and likewise, studentID’s are unique only within a university (different universities may assign the same ID to different students). Suppose we want to answer certain queries approximately from a 1/20th sample of the data. For each of the queries below, indicate how you would construct the sample. That is, tell what the key attributes should be. (a) For each university, estimate the average number of students in a course. (b) Estimate the fraction of students who have a GPA of 3.5 or more. (c) Estimate the fraction of courses where at least half the students got “A.” M a r c h 2 6 , 2 0 2 4 / D r . R S
- 25. 4.3. FILTERING STREAMS139 4.3 Filtering Streams Another common process on streams is selection, or filtering. We want to accept those tuples in the stream that meet a criterion. Accepted tuples are passed to another process as a stream, while other tuples are dropped. If the selection criterion is a property of the tuple that can be calculated (e.g., the first component is less than 10), then the selection is easy to do. The problem becomes harder when the criterion involves lookup for membership in a set. It is especially hard, when that set is too large to store in main memory. In this section, we shall discuss the technique known as “Bloom filtering” as a way to eliminate most of the tuples that do not meet the criterion. 4.3.1 A Motivating Example Again let us start with a running example that illustrates the problem and what we can do about it. Suppose we have a set S of one billion allowed email addresses – those that we will allow through because we believe them not to be spam. The stream consists of pairs: an email address and the email itself. Since the typical email address is 20 bytes or more, it is not reasonable to store S in main memory. Thus, we can either use disk accesses to determine whether or not to let through any given stream element, or we can devise a method that requires no more main memory than we have available, and yet will filter most of the undesired stream elements. Suppose for argument’s sake that we have one gigabyte of available main memory. In the technique known as Bloom filtering, we use that main memory as a bit array. In this case, we have room for eight billion bits, since one byte equals eight bits. Devise a hash function h from email addresses to eight billion buckets. Hash each member of S to a bit, and set that bit to 1. All other bits of the array remain 0. Since there are one billion members of S, approximately 1/8th of the bits will be 1. The exact fraction of bits set to 1 will be slightly less than 1/8th, because it is possible that two members of S hash to the same bit. We shall discuss the exact fraction of 1’s in Section 4.3.3. When a stream element arrives, we hash its email address. If the bit to which that email address hashes is 1, then we let the email through. But if the email address hashes to a 0, we are certain that the address is not in S, so we can drop this stream element. Unfortunately, some spam email will get through. Approximately 1/8th of the stream elements whose email address is not in S will happen to hash to a bit whose value is 1 and will be let through. Nevertheless, since the majority of emails are spam (about 80% according to some reports), eliminating 7/8th of the spam is a significant benefit. Moreover, if we want to eliminate every spam, we need only check for membership in S those good and bad emails that get through the filter. Those checks will require the use of secondary memory to access S itself. There are also other options, as we shall see when we study the general Bloom-filtering technique. As a simple example, we could use a cascade M a r c h 2 6 , 2 0 2 4 / D r . R S
- 26. 140 CHAPTER 4.MINING DATA STREAMS of filters, each of which would eliminate 7/8th of the remaining spam. 4.3.2 The Bloom Filter A Bloom filter consists of: 1. An array of n bits, initially all 0’s. 2. A collection of hash functions h1, h2, . . . , hk. Each hash function maps “key” values to n buckets, corresponding to the n bits of the bit-array. 3. A set S of m key values. The purpose of the Bloom filter is to allow through all stream elements whose keys are in S, while rejecting most of the stream elements whose keys are not in S. To initialize the bit array, begin with all bits 0. Take each key value in S and hash it using each of the k hash functions. Set to 1 each bit that is hi(K) for some hash function hi and some key value K in S. To test a key K that arrives in the stream, check that all of h1(K), h2(K), . . . , hk(K) are 1’s in the bit-array. If all are 1’s, then let the stream element through. If one or more of these bits are 0, then K could not be in S, so reject the stream element. 4.3.3 Analysis of Bloom Filtering If a key value is in S, then the element will surely pass through the Bloom filter. However, if the key value is not in S, it might still pass. We need to understand how to calculate the probability of a false positive, as a function of n, the bit-array length, m the number of members of S, and k, the number of hash functions. The model to use is throwing darts at targets. Suppose we have x targets and y darts. Any dart is equally likely to hit any target. After throwing the darts, how many targets can we expect to be hit at least once? The analysis is similar to the analysis in Section 3.4.2, and goes as follows: • The probability that a given dart will not hit a given target is (x − 1)/x. • The probability that none of the y darts will hit a given target is x−1 x y . We can write this expression as (1 − 1 x )x( y x ) . • Using the approximation (1−ǫ)1/ǫ = 1/e for small ǫ (recall Section 1.3.5), we conclude that the probability that none of the y darts hit a given target is e−y/x . M a r c h 2 6 , 2 0 2 4 / D r . R S
- 27. 4.3. FILTERING STREAMS141 Example 4.3 : Consider the running example of Section 4.3.1. We can use the above calculation to get the true expected number of 1’s in the bit array. Think of each bit as a target, and each member of S as a dart. Then the probability that a given bit will be 1 is the probability that the corresponding target will be hit by one or more darts. Since there are one billion members of S, we have y = 109 darts. As there are eight billion bits, there are x = 8 × 109 targets. Thus, the probability that a given target is not hit is e−y/x = e−1/8 and the probability that it is hit is 1 − e−1/8 . That quantity is about 0.1175. In Section 4.3.1 we suggested that 1/8 = 0.125 is a good approximation, which it is, but now we have the exact calculation. 2 We can apply the rule to the more general situation, where set S has m members, the array has n bits, and there are k hash functions. The number of targets is x = n, and the number of darts is y = km. Thus, the probability that a bit remains 0 is e−km/n . We want the fraction of 0 bits to be fairly large, or else the probability that a nonmember of S will hash at least once to a 0 becomes too small, and there are too many false positives. For example, we might choose k, the number of hash functions to be n/m or less. Then the probability of a 0 is at least e−1 or 37%. In general, the probability of a false positive is the probability of a 1 bit, which is 1 − e−km/n , raised to the kth power, i.e., (1 − e−km/n )k . Example 4.4 : In Example 4.3 we found that the fraction of 1’s in the array of our running example is 0.1175, and this fraction is also the probability of a false positive. That is, a nonmember of S will pass through the filter if it hashes to a 1, and the probability of it doing so is 0.1175. Suppose we used the same S and the same array, but used two different hash functions. This situation corresponds to throwing two billion darts at eight billion targets, and the probability that a bit remains 0 is e−1/4 . In order to be a false positive, a nonmember of S must hash twice to bits that are 1, and this probability is (1 − e−1/4 )2 , or approximately 0.0493. Thus, adding a second hash function for our running example is an improvement, reducing the false-positive rate from 0.1175 to 0.0493. 2 4.3.4 Exercises for Section 4.3 Exercise 4.3.1: For the situation of our running example (8 billion bits, 1 billion members of the set S), calculate the false-positive rate if we use three hash functions? What if we use four hash functions? ! Exercise 4.3.2: Suppose we have n bits of memory available, and our set S has m members. Instead of using k hash functions, we could divide the n bits into k arrays, and hash once to each array. As a function of n, m, and k, what is the probability of a false positive? How does it compare with using k hash functions into a single array? M a r c h 2 6 , 2 0 2 4 / D r . R S
- 28. 142 CHAPTER 4.MINING DATA STREAMS !! Exercise 4.3.3: As a function of n, the number of bits and m the number of members in the set S, what number of hash functions minimizes the false- positive rate? 4.4 Counting Distinct Elements in a Stream In this section we look at a third simple kind of processing we might want to do on a stream. As with the previous examples – sampling and filtering – it is somewhat tricky to do what we want in a reasonable amount of main memory, so we use a variety of hashing and a randomized algorithm to get approximately what we want with little space needed per stream. 4.4.1 The Count-Distinct Problem Suppose stream elements are chosen from some universal set. We would like to know how many different elements have appeared in the stream, counting either from the beginning of the stream or from some known time in the past. Example 4.5 : As a useful example of this problem, consider a Web site gath- ering statistics on how many unique users it has seen in each given month. The universal set is the set of logins for that site, and a stream element is generated each time someone logs in. This measure is appropriate for a site like Amazon, where the typical user logs in with their unique login name. A similar problem is a Web site like Google that does not require login to issue a search query, and may be able to identify users only by the IP address from which they send the query. There are about 4 billion IP addresses,2 sequences of four 8-bit bytes will serve as the universal set in this case. 2 The obvious way to solve the problem is to keep in main memory a list of all the elements seen so far in the stream. Keep them in an efficient search structure such as a hash table or search tree, so one can quickly add new elements and check whether or not the element that just arrived on the stream was already seen. As long as the number of distinct elements is not too great, this structure can fit in main memory and there is little problem obtaining an exact answer to the question how many distinct elements appear in the stream. However, if the number of distinct elements is too great, or if there are too many streams that need to be processed at once (e.g., Yahoo! wants to count the number of unique users viewing each of its pages in a month), then we cannot store the needed data in main memory. There are several options. We could use more machines, each machine handling only one or several of the streams. We could store most of the data structure in secondary memory and batch stream elements so whenever we brought a disk block to main memory there would be many tests and updates to be performed on the data in that block. Or we could use the strategy to be discussed in this section, where we 2At least that will be the case until IPv6 becomes the norm. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 29. 4.4. COUNTING DISTINCTELEMENTS IN A STREAM 143 only estimate the number of distinct elements but use much less memory than the number of distinct elements. 4.4.2 The Flajolet-Martin Algorithm It is possible to estimate the number of distinct elements by hashing the ele- ments of the universal set to a bit-string that is sufficiently long. The length of the bit-string must be sufficient that there are more possible results of the hash function than there are elements of the universal set. For example, 64 bits is sufficient to hash URL’s. We shall pick many different hash functions and hash each element of the stream using these hash functions. The important property of a hash function is that when applied to the same element, it always produces the same result. Notice that this property was also essential for the sampling technique of Section 4.2. The idea behind the Flajolet-Martin Algorithm is that the more different elements we see in the stream, the more different hash-values we shall see. As we see more different hash-values, it becomes more likely that one of these values will be “unusual.” The particular unusual property we shall exploit is that the value ends in many 0’s, although many other options exist. Whenever we apply a hash function h to a stream element a, the bit string h(a) will end in some number of 0’s, possibly none. Call this number the tail length for a and h. Let R be the maximum tail length of any a seen so far in the stream. Then we shall use estimate 2R for the number of distinct elements seen in the stream. This estimate makes intuitive sense. The probability that a given stream element a has h(a) ending in at least r 0’s is 2−r . Suppose there are m distinct elements in the stream. Then the probability that none of them has tail length at least r is (1 − 2−r )m . This sort of expression should be familiar by now. We can rewrite it as (1 − 2−r )2r m2−r . Assuming r is reasonably large, the inner expression is of the form (1 − ǫ)1/ǫ , which is approximately 1/e. Thus, the probability of not finding a stream element with as many as r 0’s at the end of its hash value is e−m2−r . We can conclude: 1. If m is much larger than 2r , then the probability that we shall find a tail of length at least r approaches 1. 2. If m is much less than 2r , then the probability of finding a tail length at least r approaches 0. We conclude from these two points that the proposed estimate of m, which is 2R (recall R is the largest tail length for any stream element) is unlikely to be either much too high or much too low. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 30. 144 CHAPTER 4.MINING DATA STREAMS 4.4.3 Combining Estimates Unfortunately, there is a trap regarding the strategy for combining the estimates of m, the number of distinct elements, that we obtain by using many different hash functions. Our first assumption would be that if we take the average of the values 2R that we get from each hash function, we shall get a value that approaches the true m, the more hash functions we use. However, that is not the case, and the reason has to do with the influence an overestimate has on the average. Consider a value of r such that 2r is much larger than m. There is some probability p that we shall discover r to be the largest number of 0’s at the end of the hash value for any of the m stream elements. Then the probability of finding r + 1 to be the largest number of 0’s instead is at least p/2. However, if we do increase by 1 the number of 0’s at the end of a hash value, the value of 2R doubles. Consequently, the contribution from each possible large R to the expected value of 2R grows as R grows, and the expected value of 2R is actually infinite.3 Another way to combine estimates is to take the median of all estimates. The median is not affected by the occasional outsized value of 2R , so the worry described above for the average should not carry over to the median. Unfortu- nately, the median suffers from another defect: it is always a power of 2. Thus, no matter how many hash functions we use, should the correct value of m be between two powers of 2, say 400, then it will be impossible to obtain a close estimate. There is a solution to the problem, however. We can combine the two methods. First, group the hash functions into small groups, and take their average. Then, take the median of the averages. It is true that an occasional outsized 2R will bias some of the groups and make them too large. However, taking the median of group averages will reduce the influence of this effect almost to nothing. Moreover, if the groups themselves are large enough, then the averages can be essentially any number, which enables us to approach the true value m as long as we use enough hash functions. In order to guarantee that any possible average can be obtained, groups should be of size at least a small multiple of log2 m. 4.4.4 Space Requirements Observe that as we read the stream it is not necessary to store the elements seen. The only thing we need to keep in main memory is one integer per hash function; this integer records the largest tail length seen so far for that hash function and any stream element. If we are processing only one stream, we could use millions of hash functions, which is far more than we need to get a 3Technically, since the hash value is a bit-string of finite length, there is no contribution to 2R for R’s that are larger than the length of the hash value. However, this effect is not enough to avoid the conclusion that the expected value of 2R is much too large. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 31. 4.5. ESTIMATING MOMENTS145 close estimate. Only if we are trying to process many streams at the same time would main memory constrain the number of hash functions we could associate with any one stream. In practice, the time it takes to compute hash values for each stream element would be the more significant limitation on the number of hash functions we use. 4.4.5 Exercises for Section 4.4 Exercise 4.4.1: Suppose our stream consists of the integers 3, 1, 4, 1, 5, 9, 2, 6, 5. Our hash functions will all be of the form h(x) = ax + b mod 32 for some a and b. You should treat the result as a 5-bit binary integer. Determine the tail length for each stream element and the resulting estimate of the number of distinct elements if the hash function is: (a) h(x) = 2x + 1 mod 32. (b) h(x) = 3x + 7 mod 32. (c) h(x) = 4x mod 32. ! Exercise 4.4.2: Do you see any problems with the choice of hash functions in Exercise 4.4.1? What advice could you give someone who was going to use a hash function of the form h(x) = ax + b mod 2k ? 4.5 Estimating Moments In this section we consider a generalization of the problem of counting distinct elements in a stream. The problem, called computing “moments,” involves the distribution of frequencies of different elements in the stream. We shall define moments of all orders and concentrate on computing second moments, from which the general algorithm for all moments is a simple extension. 4.5.1 Definition of Moments Suppose a stream consists of elements chosen from a universal set. Assume the universal set is ordered so we can speak of the ith element for any i. Let mi be the number of occurrences of the ith element for any i. Then the kth-order moment (or just kth moment) of the stream is the sum over all i of (mi)k . Example 4.6 : The 0th moment is the sum of 1 for each mi that is greater than 0.4 That is, the 0th moment is a count of the number of distinct elements in the stream. We can use the method of Section 4.4 to estimate the 0th moment of a stream. 4Technically, since mi could be 0 for some elements in the universal set, we need to make explicit in the definition of “moment” that 00 is taken to be 0. For moments 1 and above, the contribution of mi’s that are 0 is surely 0. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 32. 146 CHAPTER 4.MINING DATA STREAMS The 1st moment is the sum of the mi’s, which must be the length of the stream. Thus, first moments are especially easy to compute; just count the length of the stream seen so far. The second moment is the sum of the squares of the mi’s. It is some- times called the surprise number, since it measures how uneven the distribu- tion of elements in the stream is. To see the distinction, suppose we have a stream of length 100, in which eleven different elements appear. The most even distribution of these eleven elements would have one appearing 10 times and the other ten appearing 9 times each. In this case, the surprise number is 102 + 10 × 92 = 910. At the other extreme, one of the eleven elements could appear 90 times and the other ten appear 1 time each. Then, the surprise number would be 902 + 10 × 12 = 8110. 2 As in Section 4.4, there is no problem computing moments of any order if we can afford to keep in main memory a count for each element that appears in the stream. However, also as in that section, if we cannot afford to use that much memory, then we need to estimate the kth moment by keeping a limited number of values in main memory and computing an estimate from these values. For the case of distinct elements, each of these values were counts of the longest tail produced by a single hash function. We shall see another form of value that is useful for second and higher moments. 4.5.2 The Alon-Matias-Szegedy Algorithm for Second Moments For now, let us assume that a stream has a particular length n. We shall show how to deal with growing streams in the next section. Suppose we do not have enough space to count all the mi’s for all the elements of the stream. We can still estimate the second moment of the stream using a limited amount of space; the more space we use, the more accurate the estimate will be. We compute some number of variables. For each variable X, we store: 1. A particular element of the universal set, which we refer to as X.element, and 2. An integer X.value, which is the value of the variable. To determine the value of a variable X, we choose a position in the stream between 1 and n, uniformly and at random. Set X.element to be the element found there, and initialize X.value to 1. As we read the stream, add 1 to X.value each time we encounter another occurrence of X.element. Example 4.7 : Suppose the stream is a, b, c, b, d, a, c, d, a, b, d, c, a, a, b. The length of the stream is n = 15. Since a appears 5 times, b appears 4 times, and c and d appear three times each, the second moment for the stream is 52 +42 +32 +32 = 59. Suppose we keep three variables, X1, X2, and X3. Also, M a r c h 2 6 , 2 0 2 4 / D r . R S
- 33. 4.5. ESTIMATING MOMENTS147 assume that at “random” we pick the 3rd, 8th, and 13th positions to define these three variables. When we reach position 3, we find element c, so we set X1.element = c and X1.value = 1. Position 4 holds b, so we do not change X1. Likewise, nothing happens at positions 5 or 6. At position 7, we see c again, so we set X1.value = 2. At position 8 we find d, and so set X2.element = d and X2.value = 1. Positions 9 and 10 hold a and b, so they do not affect X1 or X2. Position 11 holds d so we set X2.value = 2, and position 12 holds c so we set X1.value = 3. At position 13, we find element a, and so set X3.element = a and X3.value = 1. Then, at position 14 we see another a and so set X3.value = 2. Position 15, with element b does not affect any of the variables, so we are done, with final values X1.value = 3 and X2.value = X3.value = 2. 2 We can derive an estimate of the second moment from any variable X. This estimate is n(2X.value − 1). Example 4.8 : Consider the three variables from Example 4.7. From X1 we derive the estimate n(2X1.value − 1) = 15 × (2 × 3 − 1) = 75. The other two variables, X2 and X3, each have value 2 at the end, so their estimates are 15 × (2 × 2 − 1) = 45. Recall that the true value of the second moment for this stream is 59. On the other hand, the average of the three estimates is 55, a fairly close approximation. 2 4.5.3 Why the Alon-Matias-Szegedy Algorithm Works We can prove that the expected value of any variable constructed as in Sec- tion 4.5.2 is the second moment of the stream from which it is constructed. Some notation will make the argument easier to follow. Let e(i) be the stream element that appears at position i in the stream, and let c(i) be the number of times element e(i) appears in the stream among positions i, i + 1, . . . , n. Example 4.9 : Consider the stream of Example 4.7. e(6) = a, since the 6th position holds a. Also, c(6) = 4, since a appears at positions 9, 13, and 14, as well as at position 6. Note that a also appears at position 1, but that fact does not contribute to c(6). 2 The expected value of n(2X.value − 1) is the average over all positions i between 1 and n of n(2c(i) − 1), that is E n(2X.value − 1) = 1 n n X i=1 n(2c(i) − 1) We can simplify the above by canceling factors 1/n and n, to get E n(2X.value − 1) = n X i=1 2c(i) − 1 M a r c h 2 6 , 2 0 2 4 / D r . R S
- 34. 148 CHAPTER 4.MINING DATA STREAMS However, to make sense of the formula, we need to change the order of summation by grouping all those positions that have the same element. For instance, concentrate on some element a that appears ma times in the stream. The term for the last position in which a appears must be 2 × 1 − 1 = 1. The term for the next-to-last position in which a appears is 2 × 2 − 1 = 3. The positions with a before that yield terms 5, 7, and so on, up to 2ma − 1, which is the term for the first position in which a appears. That is, the formula for the expected value of 2X.value − 1 can be written: E n(2X.value − 1) = X a 1 + 3 + 5 + · · · + (2ma − 1) Note that 1+3+5+· · ·+(2ma −1) = (ma)2 . The proof is an easy induction on the number of terms in the sum. Thus, E n(2X.value − 1) = P a(ma)2 , which is the definition of the second moment. 4.5.4 Higher-Order Moments We estimate kth moments, for k 2, in essentially the same way as we estimate second moments. The only thing that changes is the way we derive an estimate from a variable. In Section 4.5.2 we used the formula n(2v − 1) to turn a value v, the count of the number of occurrences of some particular stream element a, into an estimate of the second moment. Then, in Section 4.5.3 we saw why this formula works: the terms 2v − 1, for v = 1, 2, . . ., m sum to m2 , where m is the number of times a appears in the stream. Notice that 2v − 1 is the difference between v2 and (v − 1)2 . Suppose we wanted the third moment rather than the second. Then all we have to do is replace 2v−1 by v3 −(v−1)3 = 3v2 −3v+1. Then Pm v=1 3v2 −3v+1 = m3 , so we can use as our estimate of the third moment the formula n(3v2 −3v +1), where v = X.value is the value associated with some variable X. More generally, we can estimate kth moments for any k ≥ 2 by turning value v = X.value into n vk − (v − 1)k . 4.5.5 Dealing With Infinite Streams Technically, the estimate we used for second and higher moments assumes that n, the stream length, is a constant. In practice, n grows with time. That fact, by itself, doesn’t cause problems, since we store only the values of variables and multiply some function of that value by n when it is time to estimate the moment. If we count the number of stream elements seen and store this value, which only requires log n bits, then we have n available whenever we need it. A more serious problem is that we must be careful how we select the positions for the variables. If we do this selection once and for all, then as the stream gets longer, we are biased in favor of early positions, and the estimate of the moment will be too large. On the other hand, if we wait too long to pick positions, then M a r c h 2 6 , 2 0 2 4 / D r . R S
- 35. 4.5. ESTIMATING MOMENTS149 early in the stream we do not have many variables and so will get an unreliable estimate. The proper technique is to maintain as many variables as we can store at all times, and to throw some out as the stream grows. The discarded variables are replaced by new ones, in such a way that at all times, the probability of picking any one position for a variable is the same as that of picking any other position. Suppose we have space to store s variables. Then the first s positions of the stream are each picked as the position of one of the s variables. Inductively, suppose we have seen n stream elements, and the probability of any particular position being the position of a variable is uniform, that is s/n. When the (n+1)st element arrives, pick that position with probability s/(n+1). If not picked, then the s variables keep their same positions. However, if the (n + 1)st position is picked, then throw out one of the current s variables, with equal probability. Replace the one discarded by a new variable whose element is the one at position n + 1 and whose value is 1. Surely, the probability that position n + 1 is selected for a variable is what it should be: s/(n + 1). However, the probability of every other position also is s/(n + 1), as we can prove by induction on n. By the inductive hypothesis, before the arrival of the (n + 1)st stream element, this probability was s/n. With probability 1 − s/(n + 1) the (n + 1)st position will not be selected, and the probability of each of the first n positions remains s/n. However, with probability s/(n + 1), the (n + 1)st position is picked, and the probability for each of the first n positions is reduced by factor (s − 1)/s. Considering the two cases, the probability of selecting each of the first n positions is 1 − s n + 1 s n + s n + 1 s − 1 s s n This expression simplifies to 1 − s n + 1 s n + s − 1 n + 1 s n and then to 1 − s n + 1 + s − 1 n + 1 s n which in turn simplifies to n n + 1 s n = s n + 1 Thus, we have shown by induction on the stream length n that all positions have equal probability s/n of being chosen as the position of a variable. 4.5.6 Exercises for Section 4.5 Exercise 4.5.1: Compute the surprise number (second moment) for the stream 3, 1, 4, 1, 3, 4, 2, 1, 2. What is the third moment of this stream? M a r c h 2 6 , 2 0 2 4 / D r . R S
- 36. 150 CHAPTER 4.MINING DATA STREAMS A General Stream-Sampling Problem Notice that the technique described in Section 4.5.5 actually solves a more general problem. It gives us a way to maintain a sample of s stream elements so that at all times, all stream elements are equally likely to be selected for the sample. As an example of where this technique can be useful, recall that in Section 4.2 we arranged to select all the tuples of a stream having key value in a randomly selected subset. Suppose that, as time goes on, there are too many tuples associated with any one key. We can arrange to limit the number of tuples for any key K to a fixed constant s by using the technique of Section 4.5.5 whenever a new tuple for key K arrives. ! Exercise 4.5.2: If a stream has n elements, of which m are distinct, what are the minimum and maximum possible surprise number, as a function of m and n? Exercise 4.5.3: Suppose we are given the stream of Exercise 4.5.1, to which we apply the Alon-Matias-Szegedy Algorithm to estimate the surprise number. For each possible value of i, if Xi is a variable starting position i, what is the value of Xi.value? Exercise 4.5.4: Repeat Exercise 4.5.3 if the intent of the variables is to com- pute third moments. What is the value of each variable at the end? What estimate of the third moment do you get from each variable? How does the average of these estimates compare with the true value of the third moment? Exercise 4.5.5: Prove by induction on m that 1+3+5+· · ·+(2m−1) = m2 . Exercise 4.5.6: If we wanted to compute fourth moments, how would we convert X.value to an estimate of the fourth moment? 4.6 Counting Ones in a Window We now turn our attention to counting problems for streams. Suppose we have a window of length N on a binary stream. We want at all times to be able to answer queries of the form “how many 1’s are there in the last k bits?” for any k ≤ N. As in previous sections, we focus on the situation where we cannot afford to store the entire window. After showing an approximate algorithm for the binary case, we discuss how this idea can be extended to summing numbers. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 37. 4.6. COUNTING ONESIN A WINDOW 151 4.6.1 The Cost of Exact Counts To begin, suppose we want to be able to count exactly the number of 1’s in the last k bits for any k ≤ N. Then we claim it is necessary to store all N bits of the window, as any representation that used fewer than N bits could not work. In proof, suppose we have a representation that uses fewer than N bits to represent the N bits in the window. Since there are 2N sequences of N bits, but fewer than 2N representations, there must be two different bit strings w and x that have the same representation. Since w 6= x, they must differ in at least one bit. Let the last k − 1 bits of w and x agree, but let them differ on the kth bit from the right end. Example 4.10 : If w = 0101 and x = 1010, then k = 1, since scanning from the right, they first disagree at position 1. If w = 1001 and x = 0101, then k = 3, because they first disagree at the third position from the right. 2 Suppose the data representing the contents of the window is whatever se- quence of bits represents both w and x. Ask the query “how many 1’s are in the last k bits?” The query-answering algorithm will produce the same an- swer, whether the window contains w or x, because the algorithm can only see their representation. But the correct answers are surely different for these two bit-strings. Thus, we have proved that we must use at least N bits to answer queries about the last k bits for any k. In fact, we need N bits, even if the only query we can ask is “how many 1’s are in the entire window of length N?” The argument is similar to that used above. Suppose we use fewer than N bits to represent the window, and therefore we can find w, x, and k as above. It might be that w and x have the same number of 1’s, as they did in both cases of Example 4.10. However, if we follow the current window by any N − k bits, we will have a situation where the true window contents resulting from w and x are identical except for the leftmost bit, and therefore, their counts of 1’s are unequal. However, since the representations of w and x are the same, the representation of the window must still be the same if we feed the same bit sequence to these representations. Thus, we can force the answer to the query “how many 1’s in the window?” to be incorrect for one of the two possible window contents. 4.6.2 The Datar-Gionis-Indyk-Motwani Algorithm We shall present the simplest case of an algorithm called DGIM. This version of the algorithm uses O(log2 N) bits to represent a window of N bits, and allows us to estimate the number of 1’s in the window with an error of no more than 50%. Later, we shall discuss an improvement of the method that limits the error to any fraction ǫ 0, and still uses only O(log2 N) bits (although with a constant factor that grows as ǫ shrinks). To begin, each bit of the stream has a timestamp, the position in which it arrives. The first bit has timestamp 1, the second has timestamp 2, and so on. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 38. 152 CHAPTER 4.MINING DATA STREAMS Since we only need to distinguish positions within the window of length N, we shall represent timestamps modulo N, so they can be represented by log2 N bits. If we also store the total number of bits ever seen in the stream (i.e., the most recent timestamp) modulo N, then we can determine from a timestamp modulo N where in the current window the bit with that timestamp is. We divide the window into buckets,5 consisting of: 1. The timestamp of its right (most recent) end. 2. The number of 1’s in the bucket. This number must be a power of 2, and we refer to the number of 1’s as the size of the bucket. To represent a bucket, we need log2 N bits to represent the timestamp (modulo N) of its right end. To represent the number of 1’s we only need log2 log2 N bits. The reason is that we know this number i is a power of 2, say 2j , so we can represent i by coding j in binary. Since j is at most log2 N, it requires log2 log2 N bits. Thus, O(log N) bits suffice to represent a bucket. There are six rules that must be followed when representing a stream by buckets. • The right end of a bucket is always a position with a 1. • Every position with a 1 is in some bucket. • No position is in more than one bucket. • There are one or two buckets of any given size, up to some maximum size. • All sizes must be a power of 2. • Buckets cannot decrease in size as we move to the left (back in time). size 1 Two of One of size 2 At least one of size 8 Two of size 4 . . 1 0 1 1 0 1 1 0 0 0 1 0 1 1 1 0 1 1 0 0 1 0 1 1 0 1 0 1 1 0 0 0 1 0 1 1 1 0 1 1 0 0 1 . . . 1 0 1 0 1 1 0 Figure 4.2: A bit-stream divided into buckets following the DGIM rules 5Do not confuse these “buckets” with the “buckets” discussed in connection with hashing. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 39. 4.6. COUNTING ONESIN A WINDOW 153 Example 4.11 : Figure 4.2 shows a bit stream divided into buckets in a way that satisfies the DGIM rules. At the right (most recent) end we see two buckets of size 1. To its left we see one bucket of size 2. Note that this bucket covers four positions, but only two of them are 1. Proceeding left, we see two buckets of size 4, and we suggest that a bucket of size 8 exists further left. Notice that it is OK for some 0’s to lie between buckets. Also, observe from Fig. 4.2 that the buckets do not overlap; there are one or two of each size up to the largest size, and sizes only increase moving left. 2 In the next sections, we shall explain the following about the DGIM algo- rithm: 1. Why the number of buckets representing a window must be small. 2. How to estimate the number of 1’s in the last k bits for any k, with an error no greater than 50%. 3. How to maintain the DGIM conditions as new bits enter the stream. 4.6.3 Storage Requirements for the DGIM Algorithm We observed that each bucket can be represented by O(log N) bits. If the window has length N, then there are no more than N 1’s, surely. Suppose the largest bucket is of size 2j . Then j cannot exceed log2 N, or else there are more 1’s in this bucket than there are 1’s in the entire window. Thus, there are at most two buckets of all sizes from log2 N down to 1, and no buckets of larger sizes. We conclude that there are O(log N) buckets. Since each bucket can be represented in O(log N) bits, the total space required for all the buckets repre- senting a window of size N is O(log2 N). 4.6.4 Query Answering in the DGIM Algorithm Suppose we are asked how many 1’s there are in the last k bits of the window, for some 1 ≤ k ≤ N. Find the bucket b with the earliest timestamp that includes at least some of the k most recent bits. Estimate the number of 1’s to be the sum of the sizes of all the buckets to the right (more recent) than bucket b, plus half the size of b itself. Example 4.12 : Suppose the stream is that of Fig. 4.2, and k = 10. Then the query asks for the number of 1’s in the ten rightmost bits, which happen to be 0110010110. Let the current timestamp (time of the rightmost bit) be t. Then the two buckets with one 1, having timestamps t − 1 and t − 2 are completely included in the answer. The bucket of size 2, with timestamp t − 4, is also completely included. However, the rightmost bucket of size 4, with timestamp t − 8 is only partly included. We know it is the last bucket to contribute to the answer, because the next bucket to its left has timestamp less than t − 9 and M a r c h 2 6 , 2 0 2 4 / D r . R S
- 40. 154 CHAPTER 4.MINING DATA STREAMS thus is completely out of the window. On the other hand, we know the buckets to its right are completely inside the range of the query because of the existence of a bucket to their left with timestamp t − 9 or greater. Our estimate of the number of 1’s in the last ten positions is thus 6. This number is the two buckets of size 1, the bucket of size 2, and half the bucket of size 4 that is partially within range. Of course the correct answer is 5. 2 Suppose the above estimate of the answer to a query involves a bucket b of size 2j that is partially within the range of the query. Let us consider how far from the correct answer c our estimate could be. There are two cases: the estimate could be larger or smaller than c. Case 1: The estimate is less than c. In the worst case, all the 1’s of b are actually within the range of the query, so the estimate misses half bucket b, or 2j−1 1’s. But in this case, c is at least 2j ; in fact it is at least 2j+1 − 1, since there is at least one bucket of each of the sizes 2j−1 , 2j−2 , . . . , 1. We conclude that our estimate is at least 50% of c. Case 2: The estimate is greater than c. In the worst case, only the rightmost bit of bucket b is within range, and there is only one bucket of each of the sizes smaller than b. Then c = 1 + 2j−1 + 2j−2 + · · · + 1 = 2j and the estimate we give is 2j−1 + 2j−1 + 2j−2 + · · · + 1 = 2j + 2j−1 − 1. We see that the estimate is no more than 50% greater than c. 4.6.5 Maintaining the DGIM Conditions Suppose we have a window of length N properly represented by buckets that satisfy the DGIM conditions. When a new bit comes in, we may need to modify the buckets, so they continue to represent the window and continue to satisfy the DGIM conditions. First, whenever a new bit enters: • Check the leftmost (earliest) bucket. If its timestamp has now reached the current timestamp minus N, then this bucket no longer has any of its 1’s in the window. Therefore, drop it from the list of buckets. Now, we must consider whether the new bit is 0 or 1. If it is 0, then no further change to the buckets is needed. If the new bit is a 1, however, we may need to make several changes. First: • Create a new bucket with the current timestamp and size 1. If there was only one bucket of size 1, then nothing more needs to be done. However, if there are now three buckets of size 1, that is one too many. We fix this problem by combining the leftmost (earliest) two buckets of size 1. • To combine any two adjacent buckets of the same size, replace them by one bucket of twice the size. The timestamp of the new bucket is the timestamp of the rightmost (later in time) of the two buckets. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 41. 4.6. COUNTING ONESIN A WINDOW 155 Combining two buckets of size 1 may create a third bucket of size 2. If so, we combine the leftmost two buckets of size 2 into a bucket of size 4. That, in turn, may create a third bucket of size 4, and if so we combine the leftmost two into a bucket of size 8. This process may ripple through the bucket sizes, but there are at most log2 N different sizes, and the combination of two adjacent buckets of the same size only requires constant time. As a result, any new bit can be processed in O(log N) time. Example 4.13 : Suppose we start with the buckets of Fig. 4.2 and a 1 enters. First, the leftmost bucket evidently has not fallen out of the window, so we do not drop any buckets. We create a new bucket of size 1 with the current timestamp, say t. There are now three buckets of size 1, so we combine the leftmost two. They are replaced with a single bucket of size 2. Its timestamp is t − 2, the timestamp of the bucket on the right (i.e., the rightmost bucket that actually appears in Fig. 4.2. At least one of size 8 Two of size 4 size 2 Two of One of size 1 1 0 1 1 0 1 0 0 1 1 1 1 0 1 0 1 0 1 1 0 0 0 1 . . 1 0 1 Figure 4.3: Modified buckets after a new 1 arrives in the stream There are now two buckets of size 2, but that is allowed by the DGIM rules. Thus, the final sequence of buckets after the addition of the 1 is as shown in Fig. 4.3. 2 4.6.6 Reducing the Error Instead of allowing either one or two of each size bucket, suppose we allow either r − 1 or r of each of the exponentially growing sizes 1, 2, 4, . . ., for some integer r 2. In order to represent any possible number of 1’s, we must relax this condition for the buckets of size 1 and buckets of the largest size present; there may be any number, from 1 to r, of buckets of these sizes. The rule for combining buckets is essentially the same as in Section 4.6.5. If we get r + 1 buckets of size 2j , combine the leftmost two into a bucket of size 2j+1 . That may, in turn, cause there to be r + 1 buckets of size 2j+1 , and if so we continue combining buckets of larger sizes. The argument used in Section 4.6.4 can also be used here. However, because there are more buckets of smaller sizes, we can get a stronger bound on the error. We saw there that the largest relative error occurs when only one 1 from the leftmost bucket b is within the query range, and we therefore overestimate the true count. Suppose bucket b is of size 2j . Then the true count is at least M a r c h 2 6 , 2 0 2 4 / D r . R S
- 42. 156 CHAPTER 4.MINING DATA STREAMS Bucket Sizes and Ripple-Carry Adders There is a pattern to the distribution of bucket sizes as we execute the basic algorithm of Section 4.6.5. Think of two buckets of size 2j as a ”1” in position j and one bucket of size 2j as a ”0” in that position. Then as 1’s arrive in the stream, the bucket sizes after each 1 form consecutive binary integers. The occasional long sequences of bucket combinations are analogous to the occasional long rippling of carries as we go from an integer like 101111 to 110000. 1 + (r − 1)(2j−1 + 2j−2 + · · · + 1) = 1 + (r − 1)(2j − 1). The overestimate is 2j−1 − 1. Thus, the fractional error is 2j−1 − 1 1 + (r − 1)(2j − 1) No matter what j is, this fraction is upper bounded by 1/(r − 1). Thus, by picking r sufficiently large, we can limit the error to any desired ǫ 0. 4.6.7 Extensions to the Counting of Ones It is natural to ask whether we can extend the technique of this section to handle aggregations more general than counting 1’s in a binary stream. An obvious direction to look is to consider streams of integers and ask if we can estimate the sum of the last k integers for any 1 ≤ k ≤ N, where N, as usual, is the window size. It is unlikely that we can use the DGIM approach to streams containing both positive and negative integers. We could have a stream containing both very large positive integers and very large negative integers, but with a sum in the window that is very close to 0. Any imprecision in estimating the values of these large integers would have a huge effect on the estimate of the sum, and so the fractional error could be unbounded. For example, suppose we broke the stream into buckets as we have done, but represented the bucket by the sum of the integers therein, rather than the count of 1’s. If b is the bucket that is partially within the query range, it could be that b has, in its first half, very large negative integers and in its second half, equally large positive integers, with a sum of 0. If we estimate the contribution of b by half its sum, that contribution is essentially 0. But the actual contribution of that part of bucket b that is in the query range could be anything from 0 to the sum of all the positive integers. This difference could be far greater than the actual query answer, and so the estimate would be meaningless. On the other hand, some other extensions involving integers do work. Sup- pose that the stream consists of only positive integers in the range 1 to 2m for M a r c h 2 6 , 2 0 2 4 / D r . R S
- 43. 4.7. DECAYING WINDOWS157 some m. We can treat each of the m bits of each integer as if it were a separate stream. We then use the DGIM method to count the 1’s in each bit. Suppose the count of the ith bit (assuming bits count from the low-order end, starting at 0) is ci. Then the sum of the integers is m−1 X i=0 ci2i If we use the technique of Section 4.6.6 to estimate each ci with fractional error at most ǫ, then the estimate of the true sum has error at most ǫ. The worst case occurs when all the ci’s are overestimated or all are underestimated by the same fraction. 4.6.8 Exercises for Section 4.6 Exercise 4.6.1: Suppose the window is as shown in Fig. 4.2. Estimate the number of 1’s the the last k positions, for k = (a) 5 (b) 15. In each case, how far off the correct value is your estimate? ! Exercise 4.6.2: There are several ways that the bit-stream 1001011011101 could be partitioned into buckets. Find all of them. Exercise 4.6.3: Describe what happens to the buckets if three more 1’s enter the window represented by Fig. 4.3. You may assume none of the 1’s shown leave the window. 4.7 Decaying Windows We have assumed that a sliding window held a certain tail of the stream, either the most recent N elements for fixed N, or all the elements that arrived after some time in the past. Sometimes we do not want to make a sharp distinction between recent elements and those in the distant past, but want to weight the recent elements more heavily. In this section, we consider “exponentially decaying windows,” and an application where they are quite useful: finding the most common “recent” elements. 4.7.1 The Problem of Most-Common Elements Suppose we have a stream whose elements are the movie tickets purchased all over the world, with the name of the movie as part of the element. We want to keep a summary of the stream that is the most popular movies “currently.” While the notion of “currently” is imprecise, intuitively, we want to discount the popularity of a movie like Star Wars–Episode 4, which sold many tickets, but most of these were sold decades ago. On the other hand, a movie that sold M a r c h 2 6 , 2 0 2 4 / D r . R S
- 44. 158 CHAPTER 4.MINING DATA STREAMS n tickets in each of the last 10 weeks is probably more popular than a movie that sold 2n tickets last week but nothing in previous weeks. One solution would be to imagine a bit stream for each movie. The ith bit has value 1 if the ith ticket is for that movie, and 0 otherwise. Pick a window size N, which is the number of most recent tickets that would be considered in evaluating popularity. Then, use the method of Section 4.6 to estimate the number of tickets for each movie, and rank movies by their estimated counts. This technique might work for movies, because there are only thousands of movies, but it would fail if we were instead recording the popularity of items sold at Amazon, or the rate at which different Twitter-users tweet, because there are too many Amazon products and too many tweeters. Further, it only offers approximate answers. 4.7.2 Definition of the Decaying Window An alternative approach is to redefine the question so that we are not asking for a count of 1’s in a window. Rather, let us compute a smooth aggregation of all the 1’s ever seen in the stream, with decaying weights, so the further back in the stream, the less weight is given. Formally, let a stream currently consist of the elements a1, a2, . . . , at, where a1 is the first element to arrive and at is the current element. Let c be a small constant, such as 10−6 or 10−9 . Define the exponentially decaying window for this stream to be the sum t−1 X i=0 at−i(1 − c)i The effect of this definition is to spread out the weights of the stream el- ements as far back in time as the stream goes. In contrast, a fixed window with the same sum of the weights, 1/c, would put equal weight 1 on each of the most recent 1/c elements to arrive and weight 0 on all previous elements. The distinction is suggested by Fig. 4.4. Window of length 1/c Figure 4.4: A decaying window and a fixed-length window of equal weight It is much easier to adjust the sum in an exponentially decaying window than in a sliding window of fixed length. In the sliding window, we have to worry about the element that falls out of the window each time a new element arrives. That forces us to keep the exact elements along with the sum, or to use M a r c h 2 6 , 2 0 2 4 / D r . R S
- 45. 4.7. DECAYING WINDOWS159 an approximation scheme such as DGIM. However, when a new element at+1 arrives at the stream input, all we need to do is: 1. Multiply the current sum by 1 − c. 2. Add at+1. The reason this method works is that each of the previous elements has now moved one position further from the current element, so its weight is multiplied by 1 − c. Further, the weight on the current element is (1 − c)0 = 1, so adding at+1 is the correct way to include the new element’s contribution. 4.7.3 Finding the Most Popular Elements Let us return to the problem of finding the most popular movies in a stream of ticket sales.6 We shall use an exponentially decaying window with a constant c, which you might think of as 10−9 . That is, we approximate a sliding window holding the last one billion ticket sales. For each movie, we imagine a separate stream with a 1 each time a ticket for that movie appears in the stream, and a 0 each time a ticket for some other movie arrives. The decaying sum of the 1’s measures the current popularity of the movie. We imagine that the number of possible movies in the stream is huge, so we do not want to record values for the unpopular movies. Therefore, we establish a threshold, say 1/2, so that if the popularity score for a movie goes below this number, its score is dropped from the counting. For reasons that will become obvious, the threshold must be less than 1, although it can be any number less than 1. When a new ticket arrives on the stream, do the following: 1. For each movie whose score we are currently maintaining, multiply its score by (1 − c). 2. Suppose the new ticket is for movie M. If there is currently a score for M, add 1 to that score. If there is no score for M, create one and initialize it to 1. 3. If any score is below the threshold 1/2, drop that score. It may not be obvious that the number of movies whose scores are main- tained at any time is limited. However, note that the sum of all scores is 1/c. There cannot be more than 2/c movies with score of 1/2 or more, or else the sum of the scores would exceed 1/c. Thus, 2/c is a limit on the number of movies being counted at any time. Of course in practice, the ticket sales would be concentrated on only a small number of movies at any time, so the number of actively counted movies would be much less than 2/c. 6This example should be taken with a grain of salt, because, as we pointed out, there aren’t enough different movies for this technique to be essential. Imagine, if you will, that the number of movies is extremely large, so counting ticket sales of each one separately is not feasible. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 46. 160 CHAPTER 4.MINING DATA STREAMS 4.8 Summary of Chapter 4 ✦ The Stream Data Model: This model assumes data arrives at a processing engine at a rate that makes it infeasible to store everything in active storage. One strategy to dealing with streams is to maintain summaries of the streams, sufficient to answer the expected queries about the data. A second approach is to maintain a sliding window of the most recently arrived data. ✦ Sampling of Streams: To create a sample of a stream that is usable for a class of queries, we identify a set of key attributes for the stream. By hashing the key of any arriving stream element, we can use the hash value to decide consistently whether all or none of the elements with that key will become part of the sample. ✦ Bloom Filters: This technique allows us to filter streams so elements that belong to a particular set are allowed through, while most nonmembers are deleted. We use a large bit array, and several hash functions. Members of the selected set are hashed to buckets, which are bits in the array, and those bits are set to 1. To test a stream element for membership, we hash the element to a set of bits using each of the hash functions, and only accept the element if all these bits are 1. ✦ Counting Distinct Elements: To estimate the number of different elements appearing in a stream, we can hash elements to integers, interpreted as binary numbers. 2 raised to the power that is the longest sequence of 0’s seen in the hash value of any stream element is an estimate of the number of different elements. By using many hash functions and combining these estimates, first by taking averages within groups, and then taking the median of the averages, we get a reliable estimate. ✦ Moments of Streams: The kth moment of a stream is the sum of the kth powers of the counts of each element that appears at least once in the stream. The 0th moment is the number of distinct elements, and the 1st moment is the length of the stream. ✦ Estimating Second Moments: A good estimate for the second moment, or surprise number, is obtained by choosing a random position in the stream, taking twice the number of times this element appears in the stream from that position onward, subtracting 1, and multiplying by the length of the stream. Many random variables of this type can be combined like the estimates for counting the number of distinct elements, to produce a reliable estimate of the second moment. ✦ Estimating Higher Moments: The technique for second moments works for kth moments as well, as long as we replace the formula 2x − 1 (where x is the number of times the element appears at or after the selected position) by xk − (x − 1)k . M a r c h 2 6 , 2 0 2 4 / D r . R S
- 47. 4.9. REFERENCES FORCHAPTER 4 161 ✦ Estimating the Number of 1’s in a Window: We can estimate the number of 1’s in a window of 0’s and 1’s by grouping the 1’s into buckets. Each bucket has a number of 1’s that is a power of 2; there are one or two buckets of each size, and sizes never decrease as we go back in time. If we record only the position and size of the buckets, we can represent the contents of a window of size N with O(log2 N) space. ✦ Answering Queries About Numbers of 1’s: If we want to know the approx- imate numbers of 1’s in the most recent k elements of a binary stream, we find the earliest bucket B that is at least partially within the last k positions of the window and estimate the number of 1’s to be the sum of the sizes of each of the more recent buckets plus half the size of B. This estimate can never be off by more that 50% of the true count of 1’s. ✦ Closer Approximations to the Number of 1’s: By changing the rule for how many buckets of a given size can exist in the representation of a binary window, so that either r or r − 1 of a given size may exist, we can assure that the approximation to the true number of 1’s is never off by more than 1/r. ✦ Exponentially Decaying Windows: Rather than fixing a window size, we can imagine that the window consists of all the elements that ever arrived in the stream, but with the element that arrived t time units ago weighted by e−ct for some time-constant c. Doing so allows us to maintain certain summaries of an exponentially decaying window easily. For instance, the weighted sum of elements can be recomputed, when a new element arrives, by multiplying the old sum by 1 − c and then adding the new element. ✦ Maintaining Frequent Elements in an Exponentially Decaying Window: We can imagine that each item is represented by a binary stream, where 0 means the item was not the element arriving at a given time, and 1 means that it was. We can find the elements whose sum of their binary stream is at least 1/2. When a new element arrives, multiply all recorded sums by 1 minus the time constant, add 1 to the count of the item that just arrived, and delete from the record any item whose sum has fallen below 1/2. 4.9 References for Chapter 4 Many ideas associated with stream management appear in the “chronicle data model” of [8]. An early survey of research in stream-management systems is [2]. Also, [6] is a recent book on the subject of stream management. The sampling technique of Section 4.2 is from [7]. The Bloom Filter is generally attributed to [3], although essentially the same technique appeared as “superimposed codes” in [9]. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 48. 162 CHAPTER 4.MINING DATA STREAMS The algorithm for counting distinct elements is essentially that of [5], al- though the particular method we described appears in [1]. The latter is also the source for the algorithm for calculating the surprise number and higher moments. However, the technique for maintaining a uniformly chosen sample of positions in the stream is called “reservoir sampling” and comes from [10]. The technique for approximately counting 1’s in a window is from [4]. 1. N. Alon, Y. Matias, and M. Szegedy, “The space complexity of approxi- mating frequency moments,” 28th ACM Symposium on Theory of Com- puting, pp. 20–29, 1996. 2. B. Babcock, S. Babu, M. Datar, R. Motwani, and J. Widom, “Models and issues in data stream systems,” Symposium on Principles of Database Systems, pp. 1–16, 2002. 3. B.H. Bloom, “Space/time trade-offs in hash coding with allowable errors,” Comm. ACM 13:7, pp. 422–426, 1970. 4. M. Datar, A. Gionis, P. Indyk, and R. Motwani, “Maintaining stream statistics over sliding windows,” SIAM J. Computing 31, pp. 1794–1813, 2002. 5. P. Flajolet and G.N. Martin, “Probabilistic counting for database applica- tions,” 24th Symposium on Foundations of Computer Science, pp. 76–82, 1983. 6. M. Garofalakis, J. Gehrke, and R. Rastogi (editors), Data Stream Man- agement, Springer, 2009. 7. P.B. Gibbons, “Distinct sampling for highly-accurate answers to distinct values queries and event reports,” Intl. Conf. on Very Large Databases, pp. 541–550, 2001. 8. H.V. Jagadish, I.S. Mumick, and A. Silberschatz, “View maintenance issues for the chronicle data model,” Proc. ACM Symp. on Principles of Database Systems, pp. 113–124, 1995. 9. W.H. Kautz and R.C. Singleton, “Nonadaptive binary superimposed codes,” IEEE Transactions on Information Theory 10, pp. 363–377, 1964. 10. J. Vitter, “Random sampling with a reservoir,” ACM Transactions on Mathematical Software 11:1, pp. 37–57, 1985. M a r c h 2 6 , 2 0 2 4 / D r . R S
- 49. Printed Page: 1of 2 Subject Code: KIT601 0Roll No: 0 0 0 0 0 0 0 0 0 0 0 0 0 BTECH (SEM VI) THEORY EXAMINATION 2021-22 DATA ANALYTICS Time: 3 Hours Total Marks: 100 Note: Attempt all Sections. If you require any missing data, then choose suitably. SECTION A 1. Attempt all questions in brief. 2*10 = 20 Qno Questions CO (a) Discuss the need of data analytics. 1 (b) Give the classification of data. 1 (c) Define neural network. 2 (d) What is multivariate analysis? 2 (e) Give the full form of RTAP and discuss its application. 3 (f) What is the role of sampling data in a stream? 3 (g) Discuss the use of limited pass algorithm. 4 (h) What is the principle behind hierarchical clustering technique? 4 (i) List five R functions used in descriptive statistics. 5 (j) List the names of any 2 visualization tools. 5 SECTION B 2. Attempt any three of the following: 10*3 = 30 Qno Questions CO (a) Explain the process model and computation model for Big data platform. 1 (b) Explain the use and advantages of decision trees. 2 (c) Explain the architecture of data stream model. 3 (d) Illustrate the K-means algorithm in detail with its advantages. 4 (e) Differentiate between NoSQL and RDBMS databases. 5 SECTION C 3. Attempt any one part of the following: 10*1 = 10 Qno Questions CO (a) Explain the various phases of data analytics life cycle. 1 (b) Explain modern data analytics tools in detail. 1 4. Attempt any one part of the following: 10 *1 = 10 Qno Questions CO (a) Compare various types of support vector and kernel methods of data analysis. 2 (b) Given data= {2,3,4,5,6,7;1,5,3,6,7,8}. Compute the principal component using PCA algorithm. 2 M a r c h 2 6 , 2 0 2 4 / D r . R S
- 50. Printed Page: 2of 2 Subject Code: KIT601 0Roll No: 0 0 0 0 0 0 0 0 0 0 0 0 0 BTECH (SEM VI) THEORY EXAMINATION 2021-22 DATA ANALYTICS 5. Attempt any one part of the following: 10*1 = 10 Qno Questions CO (a) Explain any one algorithm to count number of distinct elements in a data stream. 3 (b) Discuss the case study of stock market predictions in detail. 3 6. Attempt any one part of the following: 10*1 = 10 Qno Questions CO (a) Differentiate between CLIQUE and ProCLUS clustering. 4 (b) A database has 5 transactions. Let min_sup=60% and min_conf=80%. TID Items_Bought T100 {M, O, N, K, E, Y} T200 {D, O, N, K, E, Y} T300 {M, A, K, E} T400 {M, U, C, K, Y} T500 {C, O, O, K, I, E} i) Find all frequent itemsets using Apriori algorithm. ii) List all the strong association rules (with support s and confidence c). 4 7. Attempt any one part of the following: 10*1 = 10 Qno Questions CO (a) Explain the HIVE architecture with its features in detail. 5 (b) Write R function to check whether the given number is prime or not. 5 M a r c h 2 6 , 2 0 2 4 / D r . R S
- 51.
- 52.
- 53. M a r c h 2 6 , 2 0 2 4 / D r . R S 11 Reference [1] https://www.jigsawacademy.com/blogs/hr-analytics/data-analytics-lifecycle/ [2]https://statacumen.com/teach/ADA1/ADA1_notes_F14.pdf [3] https://www.youtube.com/watch?v=fDRa82lxzaU [4] https://www.investopedia.com/terms/d/data-analytics.asp [5] http://egyankosh.ac.in/bitstream/123456789/10935/1/Unit-2.pdf [6] http://epgp.inflibnet.ac.in/epgpdata/uploads/epgp_content/computer_science/16._d ata_analytics/03._evolution_of_analytical_scalability/et/9280_et_3_et.pdf [7] https://bhavanakhivsara.files.wordpress.com/2018/06/data-science-and-big-data-analy -nieizv_book.pdf [8] https://www.researchgate.net/publication/317214679_Sentiment_Analysis_for_Effect ive_Stock_Market_Prediction [9] https://snscourseware.org/snscenew/files/1569681518.pdf [10] http://csis.pace.edu/ctappert/cs816-19fall/books/2015DataScienceBigDataAnalytics. pdf [11] https://www.youtube.com/watch?v=mccsmoh2_3c [12] https://mentalmodels4life.net/2015/11/18/agile-data-science-applying-kanban-in-the-a nalytics-life-cycle/ [13] https://www.sas.com/en_in/insights/big-data/what-is-big-data.html#:~:text=Big%20dat a%20refers%20to%20data,around%20for%20a%20long%20time. [14] https://www.javatpoint.com/big-data-characteristics [15] Liu, S., Wang, M., Zhan, Y., Shi, J. (2009). Daily work stress and alcohol use: Testing the cross- level moderation effects of neuroticism and job involvement. Personnel Psychology,62(3), 575–597. http://dx.doi.org/10.1111/j.1744-6570.2009.01149.x 53
- 54.