MSTMN: a novel meta-attention-based multi-task spatiotemporal network for traffic flow prediction

Traffic flow prediction in a given area is often influenced by the interactions with complex dependencies among multiple areas. By far, it remains unexplored to obtain interactive information. To address the issue, MSTMN was proposed, a multi-task learning framework that jointly learns interactive information and spatiotemporal dependencies across tasks. MSTMN consists of a node network, an edge network, and a prediction network. The node network and edge network were trained using the proposed meta-fully convolutional blocks to extract interactive features and generalizable features. The prediction network employed the meta-gated fusion and the recalibration block to both integrate these learned features and external factors. This ensures that the features capture optimal interaction information during the training phase. The proposed model was validated on two real-world movement-on-demand traffic datasets collected in Xiamen, China. Experimental results showed that MSTMN improved performance by 38.42% and 31.77% for one-step and multi-step prediction compared to the state-of-the-art baseline.