A spatio-temporal GNN and multi-agent system framework for bottleneck prediction and signal optimization in Bangalore

Urban traffic congestion remains a pressing global challenge, particularly in densely populated and rapidly urbanizing cities like Bangalore, India. This paper proposes a novel hybrid framework integrating a Spatio-Temporal Graph Neural Network (ST-GNN) and a hierarchical Multi-Agent System (MAS) for bottleneck prediction and traffic signal adjustment. This model leverages enhanced intersection-level features, edge-aware road metadata, and advanced deep learning mechanisms including GAT and ResGatedGraphConv layers. A custom Focal Loss with label smoothing is employed to handle class imbalance in bottleneck prediction. The system utilizes real-world traffic data across twelve major intersections in Bangalore, enriched with contextual factors such as weather, roadwork, and signal timing. Edge attributes are dynamically constructed using OSMnx-based topological analysis. Experimental results demonstrate that the proposed model achieves high predictive accuracy (F1-score: 0.846 at optimal threshold), and its integration with MAS enables interpretable and location-aware signal adjustment protocols. This framework presents a scalable and modular solution for traffic flow optimization in complex urban networks. • A novel GNN-MAS hybrid framework for traffic bottleneck prediction. • Application to Real-world urban traffic data from Bangalore city. • Edge-aware GNN architecture using GAT and ResGatedGraphConv. • Context-enriched node/edge features and signal adjustment policy. • High predictive accuracy with interpretable real-time recommendation.