High and Low Frequency Attention Network for Long-Term Traffic Flow Prediction

While several methods for short-term traffic prediction have been proposed in the literature, modeling the interaction between short-term and long-term traffic dynamics remains a major challenge. To address this problem, we propose a deep learning architecture named High and Low Frequency Attention Network (HLFNet) with the capability to capture multi-scale spatio-temporal dynamics. HLFNet takes inspiration from the frequency domain analysis in signal processing, where short-term (fast) and long-term (slow) dynamics are characterized in terms of high and low frequency, respectively. Specifically, we design a multi-scale attention block considering multi-grained feature information at different frequencies. The attention block comprises: a high-frequency attention operation for extracting fine-grained information within a local window; a low-frequency attention operation for extracting coarse-grained information within average pooled features. To capture the spatial interactions within the traffic network, HLFNet further considers three spatial embeddings: distance-based, attribute-based and potential-based. Experimental results with real-world traffic datasets using different prediction horizons and granularities show that, no matter how state-of-the-art methods are good at predicting in the short term, they fail to capture long-term trends. On the other hand, HLFNet retains a consistent prediction performance in both the short and the long term, outperforming all tested state-of-the-art methods in long-term traffic prediction.