An Attention-Based Bidirectional Gated Recurrent Unit Network for Location Prediction

Abstract

Locating trajectories of users has become a popular application in our daily life. How to effectively discriminate the importance of various contextual information is crucial to location prediction. Many works based on Recurrent Neural Network(RNN) models have shown promising performance by modeling the location prediction as a signal sequence processing problem and incorporating diverse characteristics into the hidden state of RNNs. However, these works rarely consider the internal relationship between locations. To overcome this limitation, we firstly propose a feedback mechanism to compensate for forward propagation, which innovatively takes advantage of the regularity of human trajectory. Then, we adopt the bidirectional Gated Recurrent Unit(GRU) network combined with data augmentation to model spatiotemporal information. Moreover, we focus on more significant factors with the attention mechanism during training, which can enhance the reliability of prediction. Experiments evaluate on three popular real-world datasets show that our model yields remarkable improvements over the compared methods.