Neurosymbolic Approach for Travel Demand Prediction: Integrating Decision Tree Rules into Neural Networks

Abstract

Accurate travel demand prediction is essential for effective transportation planning, infrastructure development, and resource optimization. Traditional models often lack both interpretability and the ability to capture complex, nonlinear relationships across geospatial and socioeconomic variables. To address this, we propose a Neurosymbolic AI framework that integrates decision tree (DT) based symbolic reasoning with neural network (NN) learning for travel demand forecasting. The model leverages multisource data including geospatial, economic, and mobility datasets and embeds interpretable if-then rules extracted from DTs as additional features in the NN. Our experiments demonstrate that this hybrid approach improves predictive accuracy across key metrics, achieving up to 24.19% reduction in Mean Absolute Error (MAE), 1.79% improvement in R², and 6.49% in Common Part of Commuters (CPC) compared to NN models. Rules extracted at finer variance thresholds (e.g., 0.0001) capture nuanced patterns and enhance model alignment with observed commuter behavior. By combining symbolic interpretability with neural generalization, the proposed method advances both the transparency and performance of travel demand modeling. The data and code can be accessed on GitHub.