Statistics-Enhanced Destination Prediction Model for Multi-Users Based on Deep Learning

Abstract

Destination prediction belongs to one of the fundamental tasks for many location-based services. However, many existing methods suffer from the data sparsity problem or poor generalization problem. To avoid these problems, this paper proposes a deep learning-based and statistics-enhanced destination prediction framework to predict the destinations of trajectories for different users simultaneously. We first analyze human mobility characteristics from a real world GPS dataset. Based on the analysis, the shared spatio-temporal regularities and user mobility preferences are combined to provide effective statistic information for multi-user destination prediction. We then propose a statistics-enhanced model where the Long Short-Term Memory (LSTM) network is the most critical component to establish complex dependencies between the trajectory sequence and statistic features. Experimental results on a realistic dataset demonstrate the significant performance improvement of the proposed framework.