A Dynamic Embedding Method for Passenger Flow Estimation

Many studies have used the embedding method to represent the traffic flow information with high dimensional embedding. Recently, due to the advancement of transfer learning technology which enhances the performance of subsequent learning tasks. The information such as locations, timestamps, and distance have been used to train a static embedding in a feature space, and the static embedding also can transfer to the subsequent task to improve performance. However, many factors affect the traffic flow prediction so more diverse traffic information needs to be considered in the pre-train embedding model. If the embedding can be dynamically obtained to generate useful features to represent a mass rapid transit (MRT) station, the features will enhance the passenger flow prediction performance of a subsequent task. Therefore, the paper proposes a dynamic pre-trained embedding model by the bidirectional encoder representations from transformers (BERT) model to represent station status and learn from traffic information in a geographical relation. To solve the problem that the fixed pre-training embedding cannot generate diversified features on different time and stations. The pre-training model also considers time and distance at the same time, and it transfers the weights of the pre-trained model to the subsequent model of passenger flow estimation for generating dynamic embedding of the station. The performance of MRT station passenger flow estimation using dynamic station embedding has significantly improved.