基于用户行为和上下文语义的分层ST-LSTM位置预测OA

Most current location prediction methods do not take user behaviour information into account.Since the user's access time and behavior pattern can reflect the function of the location,it is necessary to use the information in the pre-training of the location potential vector.In addition,when predicting the next location,the length of fine-grained sequence is too long to capture the long-distance dependence.To solve these two problems,the model Hierarchical Spatiotemporal Long Short-Term Memory Based on User Behavior and Contextual Semantics(CHST-LSTM)is proposed for location prediction.The model processes trajectory data through Transformer encoder layer,while taking user behavior into account and integrating contextual semantic of locations.Then embedding characterization of locations are obtained through pre-training.In addition,the trajectories are segmented according to the user's state,and then the form of encoder and decoder is used to extend ST-LSTM with BiLSTM used to model the context trajectory information,for calculating the long-short term dependence of the trajectories simultaneously to solve the long distance dependence problem of long sequence.Based on the analysis and experiment of the real trajectory data of the deliverymen,the unique working mode can be found by clustering.The feature vectors of locations are obtained by adding working mode and arrival time to the location embedding algorithm,and then fused to the prediction model.The results show that CHST-LSTM is more accurate in predicting the next location of the users.