MAP/REDUCE DESIGN AND IMPLEMENTATION OF APRIORIALGORITHM FOR HANDLING VOLUMINOUS DATA-SETS

Advanced Computing: An International Journal ( ACIJ ), Vol.3, No.6, November 2012
DOI : 10.5121/acij.2012.3604 29
MAP/REDUCE DESIGN AND IMPLEMENTATION OF
APRIORIALGORITHM FOR HANDLING
VOLUMINOUS DATA-SETS
Anjan K Koundinya1
,Srinath N K1
,K A K Sharma1
, Kiran Kumar1
, Madhu M N1
and Kiran U Shanbag1
1
Department of Computer Science and Engineering,
R V College of Engineering, Bangalore, India
annjank2@gmail.com
ABSTRACT
Apriori is one of the key algorithms to generate frequent itemsets. Analysing frequent itemset is a crucial
step in analysing structured data and in finding association relationship between items. This stands as an
elementary foundation to supervised learning, which encompasses classifier and feature extraction
methods. Applying this algorithm is crucial to understand the behaviour of structured data. Most of the
structured data in scientific domain are voluminous. Processing such kind of data requires state of the art
computing machines. Setting up such an infrastructure is expensive. Hence a distributed environment
such as a clustered setup is employed for tackling such scenarios. Apache Hadoop distribution is one of
the cluster frameworks in distributed environment that helps by distributing voluminous data across a
number of nodes in the framework. This paper focuses on map/reduce design and implementation of
Apriori algorithm for structured data analysis.
KEYWORDS
Frequent Itemset, Distributed Computing,Hadoop, Apriori, Distributed Data Mining
1. INTRODUCTION
In many applications of the real world, generated data is of great concern to the stakeholder as it
delivers meaningful information / knowledge that assists in making predictive analysis. This
knowledge helps in modifying certain decision parameters of the application that changes the
overall outcome of a business process. The volume of data, collectively called data-sets,
generated by the application is very large. So, there is a need of processing large data-sets
efficiently. The data-set collected may be from heterogeneous sources and may be structured or
unstructured data. Processing such data generates useful patterns from which knowledge can be
extracted. the simplest approach is to use this template and insert headings and text into it as
appropriate.
Data mining is the process of finding correlations or patterns among fields in large data-sets and
building up the knowledge-base, based on the given constraints. The overall goal of data mining
is to extract knowledge from an existing data-set and transform it into a human-understandable
structure for further use. This process is often referred to as Knowledge Discovery in data-sets
(KDD). The process has revolutionized the approach of solving the complex real-world
problems. KDD process consists of series of tasks like selection, pre-processing, transformation,
data mining and interpretation as shown in Figure1.

30
Figure 1: KDD Process
In a distributed computing environment is a bunch of loosely coupled processing
nodes connected by network. Each nodes contributes to the execution or distribution /
replication of data. It is referred to as a cluster of nodes. There are various methods of
setting up a cluster, one of which is usually referred to as cluster framework. Such
frameworks enforce the setting up processing and replication nodes for data. Examples
are DryAdLinq and Apache Hadoop (also called Map / Reduce). The other methods
involve setting up of cluster nodes on ad-hoc basis and not being bound by a rigid
framework. Such methods just involve a set of API calls basically for remote method
invocation (RMI) as a part of inter-process communication. Examples are Message
Passing Interface (MPI) and a variant of MPI called MPJExpress.
The method of setting up a cluster depends upon the data densities and up on the
scenarios listed below:
• The data is generated at various locations and needs to be accessed locally
most of the time for processing.
• The data and processing is distributed to the machines in the cluster to
reduce the impact of any particular machine being overloaded that
damages its processing
This paper is organized as follows, the next section will discuss about complete
survey of the related work carried out the domain of the distributed data mining,
specially focused on finding frequent item sets. The section 3 of this paper discusses
about the design and implementation of the Apriori algorithm tuned to the distributed
environment, keeping a key focus on the experimental test-bed requirement. The section
4, discusses about the results of the test setup based on Map/Reduce – Hadoop. Finally
conclude our work with the section 5.

31
2. RELATED WORK
Distributed Data Mining in Peer-to-Peer Networks (P2P) [1] offers an overview of
distributed data-mining applications and algorithms for peer-to-peer environments. It describes
both exact and approximate distributed data-mining algorithms that work in a decentralized
manner. It illustrates these approaches for the problem of computing and monitoring clusters in
the data residing at the different nodes of a peer-to-peer network. This paper focuses on an
emerging branch of distributed data mining called peer-to-peer data mining. It also offers a
sample of exact and approximate P2P algorithms for clustering in such distributed
environments.
Web Service-based approach for data mining in distributed environments [2] presents an
approach to develop a data mining system in distributed environments. This paper presents a
web service-based approach to solve these problems. The system is built using this approach
offers a uniform presentation and storage mechanism, platform independent interface, and a
dynamically extensible architecture. The proposed approach in this paper allows users to
classify new incoming data by selecting one of the previously learnt models.
Architecture for data mining in distributed environments [3] describes system architecture for
scalable and portable distributed data mining applications. This approach presents a document
metaphor called emph{Living Documents} for accessing and searching for digital documents
in modern distributed information systems. The paper describes a corpus linguistic analysis of
large text corpora based on colocations with the aim of extracting semantic relations from
unstructured text.
Distributed Data Mining of Large Classifier Ensembles [4] presents a new classifier
combination strategy that scales up efficiently and achieves both high predictive accuracy and
tractability of problems with high complexity. It induces a global model by learning from the
averages of the local classifiers output. The effective combination of large number of classifiers
is achieved this way.
Multi Agent-Based Distributed Data Mining [5] is the integration of multi-agent system and
distributed data mining (MADM), also known as multi-agent based distributed data mining. The
perspective here is in terms of significance, system overview, existing systems, and research
trends. This paper presents an overview of MADM systems that are prominently in use. It also
defines the common components between systems and gives a description of their strategies and
architecture.
Preserving Privacy and Sharing the Data in Distributed Environment using Cryptographic
Technique on Perturbed data [6] proposes a framework that allows systematic transformation of
original data using randomized data perturbation technique. The modified data is then submitted
to the system through cryptographic approach. This technique is applicable in distributed
environments where each data owner has his own data and wants to share this with the other
data owners. At the same time, this data owner wants to preserve the privacy of sensitive data in
the records.
Distributed anonymous data perturbation method for privacy-preserving data mining [7]
discusses a light-weight anonymous data perturbation method for efficient privacy preserving in
distributed data mining. Two protocols are proposed to address these constraints and to protect
data statistics and the randomization process against collusion attacks.An Algorithm for
Frequent Pattern Mining Based on Apriori[8] proposes three different frequent pattern mining
approaches (Record filter, Intersection and the Proposed Algorithm) based on classical Apriori
algorithm. This paper performs a comparative study of all three approaches on a data-set of

32
2000 transactions. This paper surveys the list of existing association rule mining techniques and
compares the algorithms with their modified approach.
Using Apriori-like algorithms for Spatio-Temporal Pattern Queries [9] presents a way to
construct Apriori-like algorithms for mining spatio-temporal patterns. This paper addresses
problems of the different types of comparing functions that can be used to mine frequent
patterns.Map-Reduce for Machine Learning on Multi core [10] discusses ways to develop a
broadly applicable parallel programming paradigm that is applicable to different learning
algorithms. By taking advantage of the summation form in a map-reduce framework, this paper
tries to parallelize a wide range of machine learning algorithms and achieve a significant speed-
up on a dual processor cores.
Using Spot Instances for MapReduce Work flows [11] describes new techniques to improve the
runtime of MapReducejobs. This paper presents Spot Instances (SI) as a means of attaining
performance gains at low monetary cost.
3. DESIGN AND IMPLEMENTATION
3.1 Experimental Setup
The experimental setup has three nodes connected to managed switch linked to private
LAN. One of these nodes is configured as Hadoop Master or as the namenode which
controls the data distribution over the Hadoop cluster. All the nodes are identical in
terms of the system configuration i.e., all the nodes have identical processor - Intel
Core2 Duo and assembled by standard manufacturer. As investigative
effort,configuration made to understand Hadoop will have three nodes in fully
distributed mode. The intention is to scale the number of nodes by using standard
cluster management software that can easily add new nodes to Hadoop rather than
installing Hadoop in every node. The visualization of this setup is shown in the figure 2.
Figure 2: Experimental Setup for Hadoop Multi-node
3.1.1 Special Configuration Setup for Hadoop
Focus here is on Hadoop setup with multi-node configuration. The prerequisite for
Hadoop installation is mentioned below -

33
• Set-up requires Java installed on the system, before installing Hadoop make sure
that the JDK is up and running.
• The operating system can either be Ubuntu or Suse or Fedora for the setup.
• Create a new user group and user on Linux system. This has to be replicated for
all the nodes planned for cluster setup. This user and group will exclusively
usedfor Hadoop operations.
3.1.1.1 Single Node Setup
Hadoop must first be configured to the single node setup also called as Pseudo-
distributed mode which runs on standalone machine to start with. Later this setup will
be extended to fully –configured multi-node setup with al-least two nodes and with
basic network interconnect as Cat 6 cable. Below is the required configuration setup for
single node setup-
• Download Hadoop version 0.20.x from Apache Hadoop site, which is
downloaded as tar ball (tar.gz). Change the ownership of the file as 755 or 765
(octal representation for R W E).
• Unzip the File to /usr/local and create a folder Hadoop.
• Update the bashrc file located at $HOME/.bashrc. Content to be appended to
this file is illustrated at the http://www.michael-noll.com/tutorials/running-
hadoop-on-ubuntu-linux-single-node-cluster/#update-home-bashrc.
• Modify the shell script - hadoop-env.sh and the contents to be appended is
described at the http://www.michael-noll.com/tutorials/running-hadoop-on-
ubuntu-linux-single-node-cluster/#hadoop-env-sh.
• Change the configuration file conf/core-site.xml, conf/mapred-site.xml and
conf/hdfs-site.xml with the contents illustrated at the http://www.michael-
noll.com/tutorials/running-hadoop-on-ubuntu-linux-single-node-cluster/#conf-
site-xml.
• Format the HDFS file system via Namenode by using the command given
below –
$ hduser@ubuntu: $ /usr/local/hadoop/bin/hadoopnamenode–format
• Start a single node cluster using the command -
$ hduser@ubuntu: $ /usr/local/hadoop/bin/start-all.sh
This will start up a Namenode, Datanode, Jobtracker and a Tasktracker.
• Confgure the SSH access for the user created for Hadoop on the Linux machine.
3.1.1.2 Multi-Node Configuration
• Assign names to nodes by editing the files “/etc/hosts'' . For instance -
$ sudo vi /etc/hosts

34
192.168.0.1 master
192.168.0.2 slave1
192.168.0.3 slave2
• SSH access -The hduser on master node must be able to connect to its own user
account and also hduser on slave nodes through password-less SSH login. To
connect the master to itself -
hduser@master: $ ssh master
• Configuration for Master Node -
o conf/masters -This file specifies on which machine Hadoop starts
secondary NameNodes in multi-node cluster. bin/start-dfs.sh and
bin/start-mapred.sh shell scripts will run the primary NameNode and
JobTracker on which node runs these commands respectively. Now
update the file conf/masters like this
o conf/slaves - This File actually consists of the host machines where
Hadoop slave daemonsare run. Update the file by adding these lines -
master
slave
• Modify the following configuration files previously addressed under Single node
setup-
o In the fileconf/core-site.xml
<property >
<name>fs.default.name </name>
<value>hdfs://master:54310 </value>
<description>The name of the
default file system.A URI whose schemeand authority
determine
the FileSystemimplementation.Theuri's scheme
determines the con_g property(fs.SCHEME.impl)
naming theFileSystem implementation class.Theuri's
authority is usedto determine the host,
port, etc. for a filesystem.
</description >
</property >
o In the file conf/mapred-site.xml
<property >
<name>mapred.job.tracker</name>
<value>master:54311 </value>
<description>The host and port that the MapReduce job
tracker runs at. If "local", then jobs are run in-process as a
single map and reduce task.

35
</description >
</property >
o In the file conf/hdfs-site.xml
<property >
<name>dfs.replication</name>
<value>3 </value>
<description>Default block replication. The actual number of
replications
can be specified when the file is created.The default is used if
replication
is not specified in create time.
</description >
</property >
• Format the HDFS via Namenode
/bin/start-dfs.sh
This command will brings up the HDFS with NameNode on and the DataNodes
on machines listed in /conf/slaves-file.
3.1.1.3 Running the Map/Reduce Daemon
To run mapred daemons, we need to run the following command on ?master?
hduser@master:/usr/local/hadoop$ bin/start-mapred.sh
See if all daemons have been started or not by running the command jps. i.e. on master
node
hduser@master:/usr/local/hadoop$ jps
To stop all the daemons, we have to run this command/bin/stop-all.sh.
3.1.1.4 Running the Map-Reduce job
• To run code on Hadoop cluster, put the code into a jar file and name it as
apriori.jar. Put input files in the directory /usr/hduser/input and output files in
/usr/hduser/output/.
• Now in order to run the jar file and run the following command at
HADOOP_HOME prompt hduser@master:/usr/local/hadoop$ bin/hadoop jar
apriori.jar apriori /user/hduser/input /user/hduser/output

36
3.2 System Deployment
The overall deployment of the required system is visualized using the system
organization as described in the figure 3. Series of Map calls is made to send the data to
cluster node and the format is of the form <Key, Value> and then a Reduce calls is
applied to summarize the resultant from different nodes. A simple GUI is sufficient to
display these results to user operating the Hadoop Master.
Figure 3: System Organization
3.3Algorithm Design
The algorithm discussed produces all the subsets that would be generated from
the given Item set. Further these subsets are searched against the data-sets and the
frequency is noted.There are lots of data items and their subsets, hence they need to be
searched them simultaneously so that search time reduces. Hence, the Map-Reduce
concept of the Hadoop architecture comes into picture.Map function is forked for every
subset of the items. These maps can run on any node in the distributed environment
configured under Hadoop configuration. The job distribution is taken care by the
Hadoop system and the files, data-sets required are put into HDFS. In each Map
function, the value is the item set. The whole of the data-set is scanned to find the entry
of the value item set and the frequency is noted. This is given as an output to the Reduce
function in the Reduce class defined in the Hadoop core package.
In the reducer function, each output of the each map is collected and it is put into
required file with its frequency. Algorithm is mentioned below in natural language:
• Read the subsets file.
• For each subsets in the file
o Initialise count to zero and invoke map function.
o Read the database.
o For each entry in the database
Compare with the subset.

37
If equals the increment the count by one.
Write the output counts into the intermediate file.
• Call reduce function to count the subsets and report the output to the file.
4. RESULTS
The experimental setup described before has been rigorous tested against a Pseudo-distributed
configuration of Hadoop and with standalone PC for varying intensity of data and transaction.
The fully configured multi-node Hadoop with differential system configuration (FHDSC) would
take comparatively long time to process data as against the fully configured similar multi-nodes
(FHSSC)). Similarity is in terms of the system configuration starting from computer architecture
to operating system running in it. This is clearly depicted in the figure 4.

Figure 4:FHDSC Vs. FHSSC
The results for taken from the 3-node Fully-distributed and Pseudo distributed modes of Hadoop
for large transaction are fairly good till it reaches the maximum threshold capacity of nodes. The
result is depicted in the figure 5.
Figure 5:Transactions VsHadoop Configuration

38
Looking the graph, there is large variance in time seen at threshold of 12,000 transactions.
Beyond which the time is in exponential. This is because of the computer architecture and
limited storage capacity of 80GB per Node. Hence the superset transaction generation will take
longer time to compute and the miner for frequent item-set.
The performance is expressed in the lateral comparison between FHDSC and FHSSC as given
below-
ࣁ ൌ
ࡲࡴࡰࡿ࡯
ࡲࡴࡿࡿ࡯
FHDSC = FHSSC = loge N
Where N is the number of nodes installed in the cluster.
5. CONCLUSIONS AND FUTURE ENHANCEMENTS
The paper presents a novel approach of design algorithms for clustered environment. This is
applicable to scenarios when there data-intensive computation is required. Such setup give a
broad avenue for investigation and research in Data mining. Looking the demand for such
algorithm there is urgent need to focus and explore more about clustered environment specially
for this domain.
The future works concentrates about building a framework that deals with unstructured data
analysis and complex Data mining algorithms. Also to look at integrating one of the following
parallel computing infrastructure with Map/Reduce to utilize computing power for complex
calculation on data especially in case of the classifiers under supervised learning-
• MPJ express is Java based parallel computing toolkit that can be integrated with
distributed environment.
• AtejiPX parallel programming toolkit with Hadoop.
ACKNOWLEDGEMENTS
Prof. Anjan K would like to thanks Late Dr. V.K Ananthashayana, Erstwhile Head,
Department of Computer Science and Engineering, M.S.Ramaiah Institute of
Technology, Bangalore, for igniting the passion for research.Authors would also like to
thank Late. Prof. Harish G for his consistent effort and encouragement throughout this
research project.
REFERENCES
[1]SouptikDatta, KanishkaBhaduri, Chris Giannella, Ran Wolff, and HillolKargupta, Distributed Data
Mining in Peer-to-Peer Networks, Universityof Maryland, Baltimore County, Baltimore, MD, USA,
Journal IEEEInternet Computing archive Volume 10 Issue 4, Pages 18 - 26, July 2006.
[2]Ning Chen, Nuno C. Marques, and NarasimhaBolloju, A Web Servicebasedapproach for data mining
in distributed environments, Departmentof Information Systems, City University of Hong Kong, 2005.
[3] MafruzZamanAshrafi, David Taniar, and Kate A. Smith, A Data MiningArchitecture for Distributed
Environments, pages 27-34, Springer-VerlagLondon, UK, 2007.

39
[4] GrigoriosTsoumakas and IoannisVlahavas, Distributed Data Mining ofLarge Classifier Ensembles,
SETN-2008, Thessaloniki, Greece, Proceedings,Companion Volume, pp. 249-256, 11-12 April 2008.
[5] VudaSreenivasaRao, Multi Agent-Based Distributed Data Mining: AnOver View, International
Journal of Reviews in computing, pages 83-92,2009.
[6] P.Kamakshi, A.VinayaBabu, Preserving Privacy and Sharing the Datain Distributed Environment
using Cryptographic Technique on Perturbeddata, Journal Of Computing, Volume 2, Issue 4, ISSN
21519617, April2010.
[7] Feng LI, Jin MA, Jian-hua LI, Distributed anonymous data perturbationmethod for privacy-preserving
data mining, Journal of Zhejiang UniversitySCIENCE A ISSN 1862-1775, pages 952-963, 2008.
[8] Goswami D.N. et. al., An Algorithm for Frequent Pattern Mining BasedOn Apriori (IJCSE)
International Journal on Computer Science andEngineering Vol. 02, No. 04, 942-947, 2010.
[9] MarcinGorawski and PawelJureczek, Using Apriori-like Algorithmsfor Spatio-Temporal Pattern
Queries, Silesian University of Technology,Institute of Computer Science, Akademicka 16, Poland, 2010.
[10] Cheng-Tao Chu et. al., Map-Reduce for Machine Learning on Multicore,CS Department, Stanford
University, Stanford, CA, 2006.
[11] NavrajChohanet. al., See Spot Run: Using Spot Instances for Map-Reduce Workflows, Computer
Science Department, University of California,2005.
AUTHORS PROFILE
Anjan K Koundinya has received his B.E degree from Visveswariah
Technological University, Belgaum, India in 2007 And his master degree from
Department of Computer Science and Engineering, M.S. Ramaiah Institute of
Technology, Bangalore, India. He has been awarded Best Performer PG 2010
and rank holder for his academic excellence. His areas of research includes
Network Security and Cryptology, Adhoc Networks, Mobile Computing, Agile
Software Engineering and Advanced Computing Infrastructure. He is currently
working as Assistant Professor in Dept. of Computer Science and Engineering, R
V College of Engineering.
Srinath N K has his M.E degree in Systems Engineering and Operations
Research from Roorkee University, in 1986 and PhD degree from
AvinashLingum University, India in 2009. His areas of research interests include
Operations Research, Parallel and Distributed Computing, DBMS,
Microprocessor.His is working as Professor and Head, Dept of Computer
Science and Engineering, R V College of Engineering.

MAP/REDUCE DESIGN AND IMPLEMENTATION OF APRIORIALGORITHM FOR HANDLING VOLUMINOUS DATA-SETS

More Related Content

What's hot

Similar to MAP/REDUCE DESIGN AND IMPLEMENTATION OF APRIORIALGORITHM FOR HANDLING VOLUMINOUS DATA-SETS

More from acijjournal

Recently uploaded

MAP/REDUCE DESIGN AND IMPLEMENTATION OF APRIORIALGORITHM FOR HANDLING VOLUMINOUS DATA-SETS