Multi-Perspective explainability in traffic prediction: A neuro-symbolic approach using CNN-LSTMs and concept lattices

Abstract

The rapid urbanization and continuous growth of vehicle ownership have escalated traffic congestion into a critical socioeconomic challenge, resulting in massive productivity losses and environmental degradation. To establish effective Intelligent Transportation Systems (ITS), reliable and accurate traffic forecasting is paramount. While deep learning (DL) algorithms, particularly Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN), have established themselves as state-of-the-art for spatiotemporal sequence forecasting, their inherent “black-box” nature severely limits their practical deployment in high-stakes traffic management. Decision-makers require transparent, qualitative reasoning behind algorithmic forecasts. Current post-hoc numerical Explainable AI (XAI) methods, such as SHAP and Integrated Gradients (IG), quantify feature attribution but fail to provide structured, human-readable rule mapping. To bridge the gap between predictive performance and model transparency, this paper proposes a novel Neuro-Symbolic Explainable Deep Learning framework. We integrate the statistical learning power of a hybrid CNN-LSTM architecture (the ’Neural’ component) with the rigorous logical reasoning of Formal Concept Analysis (the ’Symbolic’ component). Validated on an extensive real-world, five-year multivariate dataset from PLUS Malaysia Berhad (encompassing 73,891 aggregated records across multiple vehicle classes), the CNN-LSTM consistently outperforms standalone LSTM, BiLSTM, and GRU baselines, achieving the lowest Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). Furthermore, the synthesis of FCA with SHAP, LIME, and IG establishes a consensus view of feature importance, explicitly proving that the hybrid predictive-symbolic framework yields robust, highly accurate, and thoroughly interpretable forecasts suitable for modern ITS architectures.