Short Term Traffic Flow Prediction with Neighbor Selecting Gated Recurrent Unit

Traffic flow prediction is an important component of a modern intelligent transport system. Building an effective model for short term traffic flow prediction model is challenging. Traffic is spatial temporal in nature. A traffic flow prediction model should consider an appropriate scope of neighbourhood of traffic. To address the needs of a dynamic scope of neighbourhood. We introduce a novel gated recurrent unit variant call Neighbor Selecting Gated Recurrent Unit(NSGRU). NSGRU feature a learn-able spatial kernel with distance based K-nearest neighbor trimming scheme. Embedded external traffic knowledge are used to aid with the learning of spatial kernel. The NSGRU was evaluated with a quantized real world dataset and observed consistent improvement over baseline models.