<p>Understanding travel behavior—including travel mode, departure time, and commuting distance—is essential for effective urban transportation management and decision-making, particularly in efforts to reduce congestion and promote sustainable mobility. Despite their inherent interdependence, these commuting decisions are commonly modeled in isolation, limiting the ability to capture cross-decision relationships. This study proposes a multi-task deep learning framework that jointly predicts travel mode, departure time, and commuting distance, integrating deep neural networks with layer-wise relevance propagation to improve both predictive performance and interpretability. Using data from Beijing’s Fifth Comprehensive Traffic Survey, we demonstrate that the proposed framework consistently outperforms benchmark models, including fuzzy c-means-based support vector machines, random forests, and logistic regression, across all three commuting dimensions. Beyond accuracy gains, the model provides theory-consistent, case-based explanatory insights into the factors shaping commuting behavior. The results indicate that built environment characteristics—particularly proximity to the city center—exert a dominant influence on travel mode choice and commuting distance, whereas departure time decisions are driven primarily by individual and household attributes. In addition, the model reveals pronounced nonlinear and threshold-like relationships between key explanatory variables and travel outcomes, underscoring the importance of regional-scale spatial planning and targeted travel demand management strategies. Overall, the study highlights the value of combining multi-task learning with explainable deep learning techniques to advance both behavioral understanding and policy-relevant analysis of complex commuting decisions.</p>