Improving Destination Prediction via Ensemble of Trajectory Movement Separation and Adaptive Clustering

Destination prediction has attracted more and more attention due to its broad applications in location-based services such as traffic navigation, hotel recommendation and personalized advertising. It aims to predict future destination locations according to historical trajectory records as well as some extra information. Significant progress has been made based on Markov-chain models and neural networks. In this paper, we propose a novel Separated Trajectory Movement and Adaptive Clustering (STMAC) framework, which leverages the potential pattern of trajectory movement trend to separate historical trajectories into different categories, and then employs adaptive clustering for each category to discover fine-grained local clusters as candidate destination locations. In particular, STMAC adopts the weighted centroid of top-k discovered candidate clusters to predict a more accurate and more appropriate destination location. We conduct experiments on two real-world datasets to evaluate STMAC and other methods. Experimental results demonstrate that STMAC achieves significant improvement over competitors in terms of prediction error and accuracy. Extensive experiments on parameter analysis further show the effectiveness of our model.