Selective Trajectory Memory Network andits application in Vehicle DestinationPrediction

Abstract

Predicting efficiently the final destinations of moving vehicles can be of significant usefulness for several applications. Many probabilistic methods have been developed to address it but often include heavy feature engineering and do not generalize well to new datasets. To face these limitations, Deep-Learning models present the advantage of automating processing steps and can therefore be easily adapted to new input data. De Brébisson et al. proposed clustering based deep-learning approaches to solve it in the specific case of the prediction of Taxis destinations with remarkable performances, alongside with a proposition of a novel architecture inspired by Memory-Networks used in Natural Language Processing, and requiring no preliminary clustering. A large room for improvement was however left for the latter approach : the necessity of a relevant selection function retrieving historical trajectories similar to partial trips to predict was indeed outlined by the authors. In this work we propose to use the Segment-Path distance, introduced by Besse et al. in former works on trajectory clustering, to come up with an improved architecture of this memory model. A review of several Memory Networks architecture and their applications in time-series prediction is provided to give an overview of the different structural alternatives existing for the design of our model architecture. Finally, our model is confronted to individual car data and we propose a personalized user-by-user prediction of destinations. We discuss the suitability and limits of the type of model in this specific problem and conclude that the promising obtained results are penalized by infrequent destinations cases inducing noise whose effect could be reduced by turning our approach into a classification problem.