Dynamic Spatial-Temporal Representation Learning for Traffic Flow Prediction

As a crucial component in intelligent transportation systems, traffic flow prediction has recently attracted widespread research interest the field of artificial intelligence (AI) with increasing availability massive mobility data. Its key challenge lies how to integrate diverse factors (such as temporal rules and spatial dependencies) infer evolution trend flow. To address this problem, we propose unified neural network called Attentive Traffic Flow Machine (ATFM), which can effectively learn spatial-temporal feature representations an attention mechanism. In particular, our ATFM is composed two progressive Convolutional Long Short-Term Memory (ConvLSTM [1] ) units connected convolutional layer. Specifically, first ConvLSTM unit takes normal features input generates hidden state at each time-step, further fed into layer for map inference. The second aims learning dynamic from attentionally weighted features. Further, develop deep frameworks based on predict citywide short-term/long-term by adaptively incorporating sequential periodic data well other external influences. Extensive experiments standard benchmarks demonstrate superiority proposed method prediction. Moreover, verify generalization method, also apply customized framework forecast passenger pickup/dropoff demands show its superior performance. Our code are available https://github.com/liulingbo918/ATFM .