Riding the Rush: A Deep Learning Approach to Bus Overcrowding Prediction

Overcrowding in public transport is often explained by common-sense factors such as peak hours and busy routes. While these patterns are well known, what remains unclear is whether machine learning can uncover less obvious interactions that go beyond human intuition. This study investigates whether predictive models can not only forecast bus overcrowding but also reveal overlooked dynamics within operational data that traditional analysis might miss. Using timestamped stop-level records containing temporal and route-based features (hour of day, day of week, stop ID, line ID, and vehicle capacity), we develop machine learning and deep learning models to identify high-load events. Explainable AI (XAI) methods, including SHAP and Permutation Feature Importance, are applied to interpret the models, highlighting both the dominant drivers of crowding and the subtle combinations of factors such as weekday versus weekend shifts or route-specific vulnerabilities, that are not immediately obvious. Even with limited features, XAI also exposes blind spots, pointing to the need for richer contextual data such as weather, public events, or accessibility considerations. The implications extend beyond prediction. By revealing both known and hidden influences on bus overcrowding, the framework supports transport planners in allocating resources proactively and equitably. Vulnerable groups including disabled passengers, pregnant women, and the elderly stand to benefit from interventions that reduce barriers to safe, comfortable, and reliable travel. Ultimately, this work asks not only whether machine learning can predict overcrowding, but also whether it can help uncover what we do not yet know about when, where, and for whom it matters most.