Wavelet-Inspired Multiscale Graph Convolutional Recurrent Network for Traffic Forecasting (2401.06040v3)

Abstract: Traffic forecasting is the foundation for intelligent transportation systems. Spatiotemporal graph neural networks have demonstrated state-of-the-art performance in traffic forecasting. However, these methods do not explicitly model some of the natural characteristics in traffic data, such as the multiscale structure that encompasses spatial and temporal variations at different levels of granularity or scale. To that end, we propose a Wavelet-Inspired Graph Convolutional Recurrent Network (WavGCRN) which combines multiscale analysis (MSA)-based method with Deep Learning (DL)-based method. In WavGCRN, the traffic data is decomposed into time-frequency components with Discrete Wavelet Transformation (DWT), constructing a multi-stream input structure; then Graph Convolutional Recurrent networks (GCRNs) are employed as encoders for each stream, extracting spatiotemporal features in different scales; and finally the learnable Inversed DWT and GCRN are combined as the decoder, fusing the information from all streams for traffic metrics reconstruction and prediction. Furthermore, road-network-informed graphs and data-driven graph learning are combined to accurately capture spatial correlation. The proposed method can offer well-defined interpretability, powerful learning capability, and competitive forecasting performance on real-world traffic data sets.

• The paper introduces WavGCRN, a novel model integrating Discrete Wavelet Transform with Graph Convolutional Recurrent Networks for traffic forecasting.
• It combines predefined road-network graphs with data-driven graph learning to capture multiscale spatiotemporal traffic dynamics.
• Experiments on large-scale datasets from Los Angeles and the SF Bay Area demonstrate superior short-term prediction accuracy over existing methods.

Understanding Wavelet-Inspired Multiscale Graph Convolutional Recurrent Network for Traffic Forecasting

Introduction to Intelligent Transportation Systems

Intelligent transportation systems (ITS) are growing in prominence as cities grapple with an increase in traffic congestion challenges. One essential function at the core of ITS is traffic forecasting, which is crucial for designing smarter and more efficient transportation networks. Traditional methods for traffic prediction have relied heavily on time series analysis, requiring in-depth knowledge like queuing theory and are not always compatible with the unpredictable nature of real-world traffic data. In the wake of these traditional methods, deep learning (DL) approaches have offered a promising alternative with autonomous data analysis capabilities, minimizing the requirements for specific domain knowledge.

Advancements in Traffic Forecasting with Deep Learning

Deep learning methods harness spatiotemporal data to predict traffic, utilizing correlations inherent in the data with significant advancements being made by integration of Graph Neural Networks (GNNs) and Recurrent Neural Networks (RNNs). These models represent the state-of-the-art in traffic prediction, but certain natural characteristics of traffic data, like the multiscale structure, are not directly tackled by these models. Multiscale Analysis (MSA), an approach from mathematical and physical sciences, represents the hidden frequency information across various scales effectively, but its integration with DL into a consistent predictive model presents a challenge. This research introduces the Wavelet-Inspired Graph Convolutional Recurrent Network (WavGCRN), an innovative model designed to address this very gap.

Wavelet-Inspired Multiscale Representation in WavGCRN

WavGCRN is a unique model that integrates Discrete Wavelet Transformation (DWT) to break down traffic data into components across time and frequency, creating a multiscale representation. This process simplifies the complexity of the data, which is then processed using a multiscale encoder-decoder framework within the network. The encoder uses Graph Convolutional Recurrent Networks (GCRNs) to extract features at various scales, while a learnable Inversed DWT combined with GCRN serves as a decoder, merging the multiscale information for traffic prediction. The model also innovatively combines predefined road-network-informed graphs with data-driven graph learning to refine the representation of spatial correlations. This approach results in a model with not only high predictive performance but also a well-defined interpretability supported by the underlying physical properties of the DWT.

Experimentation and Key Findings

The WavGCRN's performance was rigorously tested on two large-scale real-world traffic datasets from Los Angeles and the San Francisco Bay Area. The model was compared with other leading traffic prediction methods, showing superior results, particularly in short-term predictions. The robustness of WavGCRN against complex traffic conditions, such as those in Los Angeles, further demonstrates its advanced predictive prowess. The findings suggest that this model could offer a new standard for traffic forecasting in intelligent transportation systems and potentially extend to other domains where time series signals are linked to spatial networks.

Conclusion of the WavGCRN Study

This paper presents a novel approach to traffic forecasting, leveraging the synergy of multiscale analysis and deep learning to create a Wavelet-Inspired Graph Convolutional Recurrent Network. The model effectively captures the multiscale spatiotemporal dynamics and improves interpretability of traffic predictions. Such advancements embody the continuous effort to modernize ITS by transforming how we process and predict traffic flows, holding the promise of more manageable and less congested future cityscapes.