DAPred: Dynamic Attention Location Prediction with Long-Short Term Movement Regularity

Predicting users’ future locations has become an important task in various aspects, such as ride-sharing, tourism recommendation and urban planning. However, existing methods disregard that users’ interest over next location is dynamic. The dynamic preference over next location involves two aspects: First, preference over distance is dynamic when users move; Second, preference over related terms vary on different target times. Hence, directly predicting next location with static network would result in unsatisfactory accuracies. Dynamic attention location prediction problem is still open now. We propose a multilayer recurrent attention model DAPred to solve the problem. The effectiveness of DAPred is underpinned by the following reasons: (1) An embedding recurrent module to map history movements into latent place, which helps build the attention module for the following layers; (2) A historical attention module that detects multiple distance preference from dynamic movement history; (3) A prediction module for learning different weights on different time gaps. Compared to the state-of-art baselines, DAPred reaches 49.8% improvement in hitting ratio accuracy, and 18.5% improvement in average distance predictor error on three real-