Intercity round-trip multi-region demand prediction based on multi-task fusion recurrent graph attention network

Existing multi-region prediction models are mainly confined to single cities and fail to fully capture the complex spatial-temporal correlations between cities. This study formally describes the problem of predicting multi-region round-trip demand between two cities and proposes a multi-task fusion recurrent graph attention network (MT-FRGAN). Firstly, this model utilizes the term frequency-inverse document frequency (TF-IDF) algorithm to measure the differences between regions and performs feature selection. Secondly, we employ a multi-head graph attention network (MGAT) to learn heterogeneous spatial correlations between adjacent regions within cities and related regions between cities, and combine it with the recurrent graph attention network (RGAN) to capture intercity multi-region temporal correlations. Lastly, we conducted multi-modal feature fusion and devised a multi-task learning mechanism to forecast demand across various types of regions. Experimental results on a dataset of round trip between Anxi County and Xiamen City demonstrate that the MT-FRGAN outperforms current single-city multi-region prediction methods by approximately 13%. Furthermore, respective ablation experiments validated the effectiveness of different components proposed in the method.