Co-attention trajectory prediction by mining heterogeneous interactive relationships

In the multi-modal spatio-temporal semantic trajectory prediction, if we can make full use of its multi-modal characteristics and heterogeneous interaction, the prediction accuracy can be significantly improved. However, the existing methods have some thorny problems. Firstly, the process of constructing an effective and semantically rich heterogeneous information interaction scene is very complex. Secondly, it is difficult to obtain interaction path instances with high quality, high relevance and high reliability. Finally, how to introduce path instances into trajectory prediction is also a difficulty. This paper proposes a common attention prediction method based on heterogeneous information network (HBCAPM). Firstly, the heterogeneous information network is constructed to make effective use of the multi-modal features in the trajectory and the heterogeneous interaction between features. Secondly, HBCAPM mines multi-source heterogeneous nodes and interaction patterns in heterogeneous information networks. Then, a path generation algorithm based on matrix decomposition and rankwalk is designed to obtain high-quality path instances. Finally, a collaborative semantic enhancement mechanism based on attention mechanism is designed to obtain the collaborative semantics of users, destinations and meta-paths. In addition, a large number of experiments on two real data sets show that HBCAPM significantly improves the effectiveness of various evaluation criteria. Compared with the latest method we discussed, the prediction accuracy is improved by 1.28% and the average distance error is reduced by 65.5 m.