Learning to Predict the Duration of an Automobile Trip

In this paper, we explore the use of machine learning and data mining to improve the prediction of travel times in an automobile. We consider two formulations of this problem, one that involves predicting speeds at different stages along the route and another that relies on direct prediction of transit time. We focus on the second formulation, which we apply to data collected from the San Diego freeway system. We report experiments on these data with k-nearest neighbout combined with a wrapper to select useful features and normalization parameters. The results suggest that 3-nearest neighbour, when using information from freeway sensors, substantially outperforms predictions available from existing digital maps. Analyses also reveal some surprises about the usefulness of other features like the time and day of the trip.