METRANS kicked off the Spring 2016 Research Seminar series with “Analysis and Prediction of Spatiotemporal Traffic Congestion,” a presentation by Dr. Ugur Demiryurek , associate director of Integrated Media Systems Center (IMSC) at USC, Viterbi. Demiryurek conducted the research with his colleague Cyrus Shahabi and doctoral student, Dingxiong Deng. 

In their research, Demiyurek et al. investigate machine learning techniques, to model and predict future traffic conditions both in normal conditions and in the case of accidents As part of their research, they crunched the numbers on the traffic delays caused by accidents in March 2015, using data gathered from over 15,000 traffic sensors throughout LA County.

The researchers were able to generate real-time predictions on the delay and traffic backlog caused by accidents.  Demiryurek demonstrated this in real-time by showing the audience their accident monitoring and prediction system, which showed an accident that had occurred just minutes before on the I-5 freeway. The system predicted that this particular accident would cause a backlog of approximately 1 mile long and would take 20 minutes to fully clear. The system also allows users to explore 4 years of historical data about where accidents and delays happen in Los Angeles most frequently.  Demiryurek mentioned that they are working on extending their algorithms to analyze and predict the impact of social events such as major sporting events, concerts and road construction zones.

Computer Science PhD Student Deng discussed the challenges of working with such large data.  Many data points were incomplete or missing entirely.  Deng also walked the audience through the project’s statistical matrices and showed how the traffic models were generated.  A machine learning algorithm compares predictions to actual conditions as they change, allowing the prediction model to constantly improve itself.

During the presentation, Dr. Christopher Lawson of the Aerospace Corporation asked how one might apply this system to larger networks, suggesting that it could be used for disaster management.  Demiryurek explained that the system is useful for drivers, who can use the predictions to get real-time information about what routes to take.  Planners can also use these numbers to better analyze transportation systems and better respond to accidents and events.

MPL student Yuzhe Ying attended the talk. “The system’s ability to catch accidents and to predict accidents’ impact is very interesting,” he said, “It provides the potential to operate traffic on a real-time level.”

Another MPL student, Jenny Hong, noted that she appreciated learning from research and interacting with researchers that she might not otherwise have the chance meet. She said that she was particularly interested in “how the researchers are able to use data gathered passively by other sources to create a model that predicts future traffic congestion and the effects of an accident.”

Ugur Demiryurek is the Associate Director  at IMSC, and is an expert on fundamental and applied data management, and machine learning fields.  His research has been funded by grants from industry partners (Microsoft Research, Oracle Labs, Intel, and HP Labs) as well as from Caltrans, METRANS, Metro, and the National Science Foundation. Previously, he worked for Fortune 500 companies in database technology development and data science positions.

Axel Hellman

Axel Hellman is a progressive degree student in his first year the Price School’s Master of Planning program.  His interests are in transit operations planning and bicycle and pedestrian planning