Selective State-Focused Graph Attention Networks for Multihorizon Traffic Forecasting

Traffic forecasting plays an important role in minimizing travel delays and enhancing road safety by using historical and real-time data to predict future traffic conditions. However, current state-of-the-art methods using graph representational learning fail to model localized complex and dynamic spatio–temporal relationships. Also, they struggle with segmenting long, context-aware traffic sequences, which are crucial for accurate forecasting. To address these issues, this work presents a novel selective state-focused graph attention network (SSGN) for traffic forecasting. SSGN utilizes a diagonally masked multihead adaptive graph attention that allows it to attend to relevant neighboring traffic nodes and analyze complex traffic patterns dynamically. Additionally, the selective state space layer is designed and integrated to model long context-aware traffic sequences to effectively capture both short and long-term sequences in traffic data. The effectiveness of SSGN is demonstrated using the simulation of urban mobility (SUMO) simulator and popular real-world benchmark datasets, namely PEMS03, PEMS04, PEMS07, and PEMS08. As a result, SSGN improved forecasting accuracy by $\sim$ 8% and $\sim$ 6% for short and long-term traffic forecasting, respectively, in comparison to other state-of-the-art methods

Attention network · Data modeling · Graph · Network traffic simulation · Traffic congestion · Traffic congestion reconstruction with Kerner's three-phase theory · Traffic generation model · Neural Networks and Applications · Time Series Analysis and Forecasting · Traffic Prediction and Management Techniques

• Spatio-Temporal Graph Convolutional Networks

  Bing Yu, Haoteng Yin et al.•Proceedings of the Twenty-Seventh…•2018

  Add to list