Self-Supervised State-Space Model for Real-Time Traffic Accident Forecasting Using eKAN Networks

Abstract

Accurate traffic accident prediction across different times and regions is vital for public safety. However, current approaches face two major challenges: (i) generalization—existing models rely heavily on manually constructed multiview structures, such as points of interest (POI) distributions and road network densities, which are difficult to scale across cities because of their labor-intensive nature; and (ii) real-Time performance—while some methods improve prediction accuracy through complex architectures, they often incur significant computational costs, hindering their real-time applicability. To address these challenges, we propose SSL-eKamba, an efficient self-supervised framework for traffic accident prediction. To improve generalization, we introduce two self-supervised auxiliary tasks that dynamically enhance traffic pattern representation by capturing spatiotemporal discrepancies. For real-time performance, we present eKamba, an optimized model based on the KAN architecture, which uses learnable univariate functions and applies a selective mechanism (selective SSM) to capture multivariate correlations. Extensive experiments on two real-world data sets demonstrate that SSL-eKamba significantly outperforms state-of-the-art baselines. Furthermore, this framework offers potential insights for other spatiotemporal tasks. The source code is available at https://github.com/KevinTan61/SSL-eKamba.