Download as PDF, PPTX
![Train the Classifier
• This is the ‘smart interface’ where GraphLab decides
which classifier to us
• Note that the features are simply flow, speed and
occupancy
• Is this model oversimplified?
33
model =
graphlab.classifier.create(final_training_data_d0_2014_1
0_29, target='incident_happened’,
features=['total_flow', 'avg_speed', 'avg_occupancy'])
Prototype the Solution](https://image.slidesharecdn.com/datasciencebasedontrafficdatafinal-150217170119-conversion-gate01/75/Get-Started-with-Data-Science-by-Analyzing-Traffic-Data-from-California-Highways-33-2048.jpg)
![SVM – Results
38
Data:
+--------------+-----------------+-------+
| target_label | predicted_label | count |
+--------------+-----------------+-------+
| 0 | 0 | 3746 |
| 0 | 1 | 12330 |
| 1 | 0 | 3522 |
| 1 | 1 | 3711 |
+--------------+-----------------+-------+
[4 rows x 3 columns]
, 'accuracy': 0.31991934445922177}
Prototype the Solution](https://image.slidesharecdn.com/datasciencebasedontrafficdatafinal-150217170119-conversion-gate01/75/Get-Started-with-Data-Science-by-Analyzing-Traffic-Data-from-California-Highways-38-2048.jpg)
![Boosted Trees – Results
39
Data:
+--------------+-----------------+-------+
| target_label | predicted_label | count |
+--------------+-----------------+-------+
| 0 | 0 | 15082 |
| 0 | 1 | 994 |
| 1 | 0 | 5574 |
| 1 | 1 | 1659 |
+--------------+-----------------+-------+
[4 rows x 3 columns]
, 'accuracy': 0.7182204298768716}
Prototype the Solution](https://image.slidesharecdn.com/datasciencebasedontrafficdatafinal-150217170119-conversion-gate01/75/Get-Started-with-Data-Science-by-Analyzing-Traffic-Data-from-California-Highways-39-2048.jpg)
![Train the Classifier
• This is the ‘smart interface’ where GraphLab decides
which classifier to us
• Note that the features are simply flow, speed and
occupancy
• Is this model oversimplified?
33
model =
graphlab.classifier.create(final_training_data_d0_2014_1
0_29, target='incident_happened’,
features=['total_flow', 'avg_speed', 'avg_occupancy'])
Prototype the Solution](https://crownmelresort.com/image.slidesharecdn.com/datasciencebasedontrafficdatafinal-150217170119-conversion-gate01/75/Get-Started-with-Data-Science-by-Analyzing-Traffic-Data-from-California-Highways-33-2048.jpg)
![SVM – Results
38
Data:
+--------------+-----------------+-------+
| target_label | predicted_label | count |
+--------------+-----------------+-------+
| 0 | 0 | 3746 |
| 0 | 1 | 12330 |
| 1 | 0 | 3522 |
| 1 | 1 | 3711 |
+--------------+-----------------+-------+
[4 rows x 3 columns]
, 'accuracy': 0.31991934445922177}
Prototype the Solution](https://crownmelresort.com/image.slidesharecdn.com/datasciencebasedontrafficdatafinal-150217170119-conversion-gate01/75/Get-Started-with-Data-Science-by-Analyzing-Traffic-Data-from-California-Highways-38-2048.jpg)
![Boosted Trees – Results
39
Data:
+--------------+-----------------+-------+
| target_label | predicted_label | count |
+--------------+-----------------+-------+
| 0 | 0 | 15082 |
| 0 | 1 | 994 |
| 1 | 0 | 5574 |
| 1 | 1 | 1659 |
+--------------+-----------------+-------+
[4 rows x 3 columns]
, 'accuracy': 0.7182204298768716}
Prototype the Solution](https://crownmelresort.com/image.slidesharecdn.com/datasciencebasedontrafficdatafinal-150217170119-conversion-gate01/75/Get-Started-with-Data-Science-by-Analyzing-Traffic-Data-from-California-Highways-39-2048.jpg)
Uploaded byAerospike, Inc.
Get Started with Data Science by Analyzing Traffic Data from California Highways
This document summarizes an effort to analyze traffic data from California highways to better understand data science techniques. The researchers searched for an open dataset, eventually finding sensor data from California highways. They analyzed the data format and values to understand it. To detect traffic incidents, they framed it as a classification problem and prepared training data by labeling sensor records near incidents as positive examples. They trained classifiers on this data but initial results were poor. After refining the features and balancing the training data, the classifiers showed more promising results.
More Related Content
Similar to Get Started with Data Science by Analyzing Traffic Data from California Highways
Get Started with Data Science by Analyzing Traffic Data from California Highways
- 1. GET STARTED WITHDATA SCIENCE BYANALYZING TRAFFIC DATA FROM CALIFORNIA HIGHWAYS Cirrus Shakeri, Ph.D. Inventurist LLC 1 With contribution and support from: • Aerospike Team: Brian Bulkowski, Monica Pal, Dash Desai, and Ondrej Jaura • GraphLab Team: Danny Bickson and Eduardo Rosini
- 2. The Short Story • What we wanted to achieve • To understand how data science works and what results can be achieved • How repeatable and reusable the data science techniques can be • Broaden the use of Aerospike in Realtime Analytics • The starting point: data • Hard to find the right dataset • Datasets from California highways • Application example: realtime detection of traffic incidents (a hypothesis) • The non-glamorous part of data science • Understanding the data and the domain in reasonable depth • Data cleansing and preparation • The glamorous part • The algorithms– can we treat machine learning as a black box? • The reality of Data Science hits: We need a data scientist! • Glimpses of hope • Some promising results but the hard work still remains—please join the effort 2
- 3. THE LONGER STORY Ourlessons learned and how you can apply these techniques in your applications 3
- 4.
- 5. Search for theright dataset • Criteria • Open, realtime, interesting, not-a-toy dataset, and not-creepy! • California highways sensor data • 16,490 stations, 41,470 sensors across California highways 5 Courtesy of California Department of Transportation Start from Data
- 6. Checklist for understandingthe data • Download a sample and study it • what are the attributes, what is the range of values, … • How is data structured? • tables, documents, graph, etc. • How do we access it? • API, file downloads, … • How is the quality of data? • missing values, incorrect values, … • How is data generated? • How fast? Do we have access to the historical data? How much of it? • Are there legal and ethical issues in analyzing the data • privacy, security, … 6Understand the Dataset
- 7. Data format andstructure “d04_text_station_raw_2014_10_29.txt” 7 CSV format 8,403,282 lines 565.5 MB Understand the Dataset
- 8. What exactly isin the data 8 station ID Average speed for lane 1: 44 miles per hour lane 2 lane 3 lane 4 lane 5 vehicle flow for lane 1: 17 vehicles passed in 30 sec on lane 1 average occupancy for lane 1 lane 1 was occupied 18% of time in 30 sec Understand the Dataset
- 9. Deeper study ofthe data • GraphLab + iPython Notebook • Both open source and free 9 www.dato.com ipython.org/notebook.html Pick Tools
- 10. Load Sample Data 10 graphlab.SFrame.read_csv(url_d04_text_station_raw_2014_10_29,...) ... PROGRESS: Unable to parse line "10/29/2014 00:44:36,400728,2,... PROGRESS: Unable to parse line "10/29/2014 00:00:05,400402,2,... PROGRESS: Unable to parse line "10/29/2014 01:29:57,401542,1,... PROGRESS: Unable to parse line "10/29/2014 00:44:36,401211,0,... PROGRESS: Read 778502 lines. Lines per second: 174789 PROGRESS: Read 3707451 lines. Lines per second: 340087 PROGRESS: Read 6514640 lines. Lines per second: 377548 PROGRESS: 577297 lines failed to parse correctly PROGRESS: Finished parsing file /Users/cirrus/Documents/Aerospike/... PROGRESS: Parsing completed. Parsed 7825984 lines in 19.8226 secs. Understand the Dataset
- 11. Study Range ofValues in the Dataset 11 ! Understand the Dataset
- 12. Lesson Learned: Needa fast database • Exploring sample files worked as a start • Next we needed to slice and dice a large amounts of the traffic data in various ways • How is speed at a specific station changes over the course of a year at a specific time of day? • Are there repeatable seasonal or weekly patterns? • What about traffic behavior along a specific highway, say US-101 in Bay Area? • ... • This requires loading a large amount of historical data in a database and running queries • Exploring data should be interactive so the database must be fast • Current plan • Load a good portion of the highway traffic data into Aerospike • Provide developer community free API access • Experiment with the data and collaborative on building applications • The motivation is to open up the use of the Aerospike technology to different applications especially in realtime analytics 12Pick Tools
- 13. Good Data vs.Bad Data 13Check the Data Quality
- 14. Imputed Data: 5-minaggregate 14Check the Data Quality
- 15. Two more datasets • Station locations • Incident data 15Add More Datasets
- 16. Detecting Incidents inRealtime • Can we detect accidents based on sensor data? • Is there a value in such a prediction? • Can we detect accidents faster than the way they are reported now? • Is it doable? Has been done before? • Can we ‘predict’ accidents? • A better understanding of the domain is needed 16Generate Application Ideas
- 17.
- 18. Literature Survey • Typesof incidents: Collision, Debris, Break-down 18Study the Domain Reference: “REAL-TIME DETECTION OF ROAD TRAFFIC INCIDENTS”, by PERO ŠKORPUT, et al
- 19. Literature Survey 19Study theDomain Reference: “REAL-TIME DETECTION OF ROAD TRAFFIC INCIDENTS”, by PERO ŠKORPUT, et al
- 20. Can incidents bedetected in near realtime? (under 1 minute) 20 “The survey responses point to a general consensus that the unacceptably high rates of false alarms generated by available incident detection algorithms is the major deterrent.” Study the Domain
- 21. Algorithmic Approaches • Comparativealgorithms • Current traffic data is compared with the ‘normal’ threshold • Statistical algorithms • Current traffic data is compared with forecasted data • Traffic theory based algorithms • Rapid change in speed while flow or occupancy do not change considerably (based on ‘catastrophe theory’) • Advanced algorithms • AI and neural networks 21Study the Domain Reference: “ OVERVIEW TO SOME INCIDENT DETECTION ALGORITHMS: A COMPARATIVE EVALUATION WITH ISTANBUL FREEWAY DATA” by Onur Deniz, et al
- 22.
- 23. Study Patterns andCorrelations 23Explore the Data
- 24. Exploring the Data– Speed 24Explore the Data
- 25. Exploring the Data– Occupancy 25Explore the Data
- 26. Exploring the Data– Flow 26Explore the Data
- 27. FORMULATE THE DATA SCIENCEAPPROACH 27
- 28. Detecting Incidents asa Classification Problem • Features • Started simple: Flow, occupancy and speed • Next: prepare training data 28 Any individual record classified as incident or no-incident: Prototype the Solution
- 29. Classifiers in GraphLab • Classification models supported in GraphLab • support vector machines (SVM) • classify instances based on a linear function of the features • logistic regression • estimates the probability of instances belonging to a class as a logistic function of a linear combination of features • boosted trees • combines the results of a set of base classification decision trees • neural networks • middle layers of the network compute composite, intermediary features • Also, a smart interface that selects the right model based on the data 29Prototype the Solution
- 30. PREPARE TRAINING DATA TRAINTHE MODEL EVALUATE THE MODEL 30
- 31. Why do weneed training data? • Supervised learning: a classifier is trained via a dataset that is labeled with positive and negative instances 31 downstream occupancy / speed upstreamoccupancy/speed incident no-incident ? Prototype the Solution
- 32. Training Data –the end result 32Prototype the Solution
- 33. Train the Classifier • This is the ‘smart interface’ where GraphLab decides which classifier to us • Note that the features are simply flow, speed and occupancy • Is this model oversimplified? 33 model = graphlab.classifier.create(final_training_data_d0_2014_1 0_29, target='incident_happened’, features=['total_flow', 'avg_speed', 'avg_occupancy']) Prototype the Solution
- 34. First results wereembarrassing! • Basically no incident could be detected in the test data • The reality of data science hit! • Leveraged the GraphLab on-line community • Worked with GraphLab Data Scientist: Danny Bickson • Made some changes • Added more features • Station id, Num of lanes, Direction of travel, Absolute postmile, Hour/ Min/Day that incident occurred • Balanced the negative and positive examples • Tagged station data in a 3-mile vicinity of an incident as positive instances • Created classifiers for specific highways (e.g., US-101) • … 34Prototype the Solution
- 35. Training Data PreparationSteps • Load CSV files for 5-min aggregate sensor data • From Oct 29 – Nov 4, 2014 • In District 4 (Bay Area) • Prepare sensor dataset • Filter for highway US-101 • 1,861,626 records • Convert data types • String to datatime • Split timestamp and add as separate features • Add day of week as extra feature • Drop records with unknown values • 1,447,195 records 35Prototype the Solution
- 36. Training Data PreparationSteps • Load CSV file for sensor locations • 920 sensor locations for US101S in District 4 • Convert highway data to match sensor data • SR101S à highway = 101, direction of travel = ‘S’ • Join with sensor data • Load CSV file for incidents from Oct 29 – Nov 4 • Covert datatypes, drop records with unknown values, … • Find sections of highways 101 in District 4 • Filter for incident data for 101S in District 4 • 176 incidents • Create a nearest neighbor model for sensor locations • To query the nearest sensors to an incident 36Prototype the Solution
- 37. Training Data PreparationSteps • For each incident on highway 101S in Bay Area between Oct 29 and Nov 4: • Find the sensor stations in the 3-miles radius • For each sensor in the 3 miles radius • Extract the sensor data for any timestamp that falls within the ‘duration of the incident’ • Add sensor data to the ‘positive examples’ in the training data • 28,460 records • Note: upstream/downstream stations not taken into account • ‘Negative Examples’ data (no incidents) is un-proportionally larger (10 times) • Extract a 10% sample of the negative examples data • 24,026 records • Add sampled sensor data for negative examples to the training data • 52,486 records 37Prototype the Solution
- 38. SVM – Results 38 Data: +--------------+-----------------+-------+ |target_label | predicted_label | count | +--------------+-----------------+-------+ | 0 | 0 | 3746 | | 0 | 1 | 12330 | | 1 | 0 | 3522 | | 1 | 1 | 3711 | +--------------+-----------------+-------+ [4 rows x 3 columns] , 'accuracy': 0.31991934445922177} Prototype the Solution
- 39. Boosted Trees –Results 39 Data: +--------------+-----------------+-------+ | target_label | predicted_label | count | +--------------+-----------------+-------+ | 0 | 0 | 15082 | | 0 | 1 | 994 | | 1 | 0 | 5574 | | 1 | 1 | 1659 | +--------------+-----------------+-------+ [4 rows x 3 columns] , 'accuracy': 0.7182204298768716} Prototype the Solution
- 40. Ideas for furtherimproving the results • Work with ‘raw’ station data instead of 5-min aggregates • Further explore the data for entire 2014 and entire California highways • It’s not easy when you are dealing with files: load into Aerospike to rescue • Add more features • difference between values in neighboring stations • difference between consecutive times at the same location • ... • Try different models for rush-hour and off-peak hours • Take into account seasonal effects • Validate the actual time of incident in training data • Include upstream and downstream data as separate features • Investigate causality • Is an incident causing slow-down or is it the other way around? • Stretch goal: Can we ‘predict’ incidents before happening? • Leading indicators for accidents • Improve safety of self-driving cars 40Prototype the Solution
- 41.
- 42. Data Science –it’s iterative 42 Data Domain Problem Data Science Prototype Product Last Word
- 43.
- 44. Take Away 1:Realtime Analytics • Realtime (fast) vs. offline analytics (slow) • Getting closer to how human brain makes decisions The next big thing: • The line between capturing and analyzing data is disappearing • And technologies such as Aerospike and GraphLab are helping! 44Last Word
- 45. Take Away 2:Where does all this lead? 1. More intelligence for us • ‘Augmenting Human Intellect’ • Doug Engelbart 2. More automation for machines • ‘The Second Machine Age’ • Erik Brynjolfsson, Andrew McAfee 45Last Word
- 46. THANK YOU Cirrus Shakeri,Ph.D. Co-Founder, Inventurist +1 650-380-9794 twitter.com/cirrus_shakeri www.linkedin.com/in/cshakeri/ cirrus.shakeri@inventurist.com 46
- 47. Location (mile) (Abs PM) Time 00:00:00 23:55:00 0.05.0 10.0 Incident Incident start time Incident location Station location 5-min aggregate end timestamp 5 min Distance bet stations 5-min aggregate start timestamp Visualize an incident at a time and location 47