MDAN: Multimodal Double-Layer Attention Network for Next-Location Prediction

With the rapid development of social media, a large amount of user check-in and trajectory data has been generated. Using these trajectory data to analyze user movement patterns is of great value to urban traffic planning, infrastructure configuration, and third-party services. The existing trajectory prediction models have the following drawbacks: 1) the models of recurrent neural network(RNN) generate forgetfulness when performing long trajectory sequence prediction; 2) the ability to model the multi-level periodicity of user movement is not well. In this paper, we propose MDAN, a multimodal double-layer attention network to predict the next location. It takes user, location, time, latitude, and longitude as the input feature vectors. We design a multi-headed self-attentive mechanism to predict long trajectory sequences while capturing the multi-level periodic patterns of users. Based on the idea that the next moving location of the user is likely to be in the region of the current location, we derive the region features of each location by using the region division method. Then it matches the feature vectors of the user's movement patterns by the linear attention mechanism. We obtain the next most likely location of the user finally. The experimental results show that our model achieves the prediction of long trajectory sequences on the Foursquare-NYC dataset. And the top1 accuracy on this dataset is 18.23% higher than the DeepMove model, which proves the effectiveness of our model.