Neuro‐Symbolic TabNet for Explainable Congestion Management in Vehicular Ad Hoc Networks (VANETs)

ABSTRACT

Vehicular ad hoc networks (VANETs) play a crucial role in modern intelligent transportation systems, yet most learning‐based congestion controllers operate as black boxes, limiting trust, auditability and deployment in safety‐critical settings. This work addresses this gap by proposing a neuro‐symbolic framework that couples a TabNet‐based congestion classifier with an explicit symbolic rule layer to provide both accurate and interpretable traffic‐management decisions. Using the VANET Traffic Congestion Dataset, we preprocess heterogeneous traffic, communication and environmental features, train an enhanced TabNet model with contextual embeddings and distil its behaviour into a compact set of human‐readable rules with quantified support, coverage, confidence and fidelity. The framework is evaluated through statistical validation, ablation studies and an observed post‐decision queue analysis that compares recorded queue lengths before and after the recommended decision intervals. Results show accuracy and macro‐F1 scores around 0.99, micro‐average AUC near 0.996, high aggregate surrogate fidelity of 99.2% and more than 60% of evaluated scenarios exhibiting reduced queues, all while maintaining millisecond‐level inference latency suitable for VANET edge devices. These findings demonstrate that the proposed neuro‐symbolic approach can match state‐of‐the‐art predictive performance while yielding transparent, rule‐based explanations and post‐decision queue observations linked to recommended decisions. The framework therefore offers a practical path toward trustworthy, human‐centric AI for real‐time vehicular communication and traffic control.