Predicting Travel Times with Context-Dependent Random Forests by Modeling Local and Aggregate Traffic Flow

Predicting future traffic flow has the potential to decrease travel times. This paper describes a methodology developed to predict future travel times based on GPS reports of 1% of the cars on the road in a simulation framework. Features representing local and aggregate models of traffic flow were extracted from these reports. These included the counts of moving and stopped cars in different areas, along with each road segment's mean speed. Context-dependent Random Forests were then trained to predict traffic flow six and thirty minutes into the future. This algorithm performed best in one track of the 2010 IEEE ICDM Contest: TomTom Traffic Prediction for Intelligent GPS Navigation, improving 9% on the next-best algorithm and 62.5% on the baseline.