A multi-task memory network with knowledge adaptation for multimodal demand forecasting

Travel demand forecasting is useful for both trip and service planning, and thus is of great importance. Most existing studies focus on demand forecasting for a single mode, while much less attention has been paid to multimodal demand forecasting. This paper develops a multimodal demand forecasting approach, which can learn and utilize information/knowledge from different public transit modes and thus improve the demand prediction of the travel mode with sparse observations (e.g., station-sparse mode). In particular, this study focuses on improving the passenger demand prediction accuracy of the station-sparse mode(s) with the help of the station-intensive mode (i.e., the mode with more sufficient knowledge and intensive station distribution over space). We propose a novel Knowledge Adaptation with Attentive Multi-task Memory Network (KA2M2) in order to utilize closely-related demand patterns from the station-intensive mode for demand forecasting of the station-sparse mode(s). Specifically, we first design a memory-augmented recurrent network for enhancing the ability to capture the long-and-short term demand information and storing the extracted temporal knowledge of each transit mode. Then, we develop and integrate an attention-based knowledge adaptation module to adapt relevant information from the station-intensive source to the station-sparse source(s). The experimental results on a real-world dataset collected from the Greater Sydney area covering four public transport modes (bus, train, light rail, and ferry) demonstrate that the proposed approach consistently outperforms a number of baseline methods and state-of-the-art models. Our findings also illustrate that incorporating information/knowledge from multimodal trip records can enhance the demand forecasting accuracy for station-sparse modes.