Short-Term Traffic Forecasting: Modeling and Learning Spatio-Temporal Relations in Transportation Networks Using Graph Neural Networks

Behrooz Shahsavari and Pieter Abbeel

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2015-243

December 17, 2015

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-243.pdf

This report addresses the problem of traffic conditions forecasting based on big data with a novel artificial intelligence approach. We propose an empirical data-driven graph-oriented model that can process spatio-temporal measurements. The proposed intelligent model not only incorporates observations from multiple locations, but also, explicitly takes into account the spatial interrelations (between the sensor locations) that are forced by the traffic network topology. We abstract the data collected in a transportation network by a graph that has a set of “nodes” corresponding to the sensor locations and a set of “edges” representing the spatial interrelations governed by the network topology. Both entities are associated with real valued feature vectors. For instance, a history of traffic flow, occupancy and speed measured by a sensor form the feature vector associated with the corresponding node. On the other hand, the road characteristics such as length, capacity and direction constitute the edge feature vectors. A Graph Neural Network is trained in a supervised fashion to predict future traffic conditions based on the stated graph-structured data. This model combines the advantages of methods like Cell Transmission Model that benefits from knowing causalities enforced by traffic network topology, and advantages of neural network models that can extract very complex and nonlinear relations after being trained on big data. Moreover, the proposed model is _robust_ to sample missing. A comprehensive empirical study is conducted on traffic data obtained from PeMS database. A method is proposed to constitute the spatial interrelations (i.e. graph edges) between the sensor locations by deploying Google Maps (Directions) API. We evaluate the effectiveness of the proposed prediction method in locations with simple and complex dynamics (e.g. the intersection of several highways with multitude on_ and off-ramps) individually.

Advisors: Pieter Abbeel

BibTeX citation:

@mastersthesis{Shahsavari:EECS-2015-243,
    Author= {Shahsavari, Behrooz and Abbeel, Pieter},
    Title= {Short-Term Traffic Forecasting: Modeling and Learning Spatio-Temporal Relations in Transportation Networks Using Graph Neural Networks},
    School= {EECS Department, University of California, Berkeley},
    Year= {2015},
    Month= {Dec},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-243.html},
    Number= {UCB/EECS-2015-243},
    Abstract= {This report addresses the problem of traffic conditions forecasting based on big data with a novel artificial intelligence approach. We propose an empirical data-driven graph-oriented model that can process spatio-temporal measurements. The proposed intelligent model not only incorporates observations from multiple locations, but also, explicitly takes into account the spatial interrelations (between the sensor locations) that are forced by the traffic network topology. We abstract the data collected in a transportation network by a graph that has a set of “nodes” corresponding to the sensor locations and a set of “edges” representing the spatial interrelations governed by the network topology. Both entities are associated with real valued feature vectors. For instance, a history of traffic flow, occupancy and speed measured by a sensor form the feature vector associated with the corresponding node. On the other hand, the road characteristics such as length, capacity and direction constitute the edge feature vectors. A Graph Neural Network is trained in a supervised fashion to predict future traffic conditions based on the stated graph-structured data.
This model combines the advantages of methods like Cell Transmission Model that benefits from knowing causalities enforced by traffic network topology, and advantages of neural network models that can extract very complex and nonlinear relations after being trained on big data. Moreover, the proposed model is _robust_ to sample missing.
A comprehensive empirical study is conducted on traffic data obtained from PeMS database. A method is proposed to constitute the spatial interrelations (i.e. graph edges) between the sensor locations by deploying Google Maps (Directions) API. We evaluate the effectiveness of the proposed prediction method in locations with simple and complex dynamics (e.g. the intersection of several highways with multitude on_ and off-ramps) individually.},
}

EndNote citation:

%0 Thesis
%A Shahsavari, Behrooz 
%A Abbeel, Pieter 
%T Short-Term Traffic Forecasting: Modeling and Learning Spatio-Temporal Relations in Transportation Networks Using Graph Neural Networks
%I EECS Department, University of California, Berkeley
%D 2015
%8 December 17
%@ UCB/EECS-2015-243
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-243.html
%F Shahsavari:EECS-2015-243