Addressing Imbalanced Records for Accurate Hourly Boarding Demand Prediction in Buses

Smart-card tap-on data provides essential insights into passenger boarding patterns and aids in forecasting bus travel demand. However, these datasets often suffer from class imbalance, where instances of actual boarding events at specific stops and times are much fewer compared to non-boarding events. This imbalance negatively affects the accuracy of predictive models designed to estimate hourly boarding volumes. To tackle this challenge, this study employs deep generative adversarial networks (Deep-GAN) to synthesize realistic boarding instances, creating a more balanced training dataset. This enhanced dataset is then utilized to train a deep neural network (DNN) for predicting boarding events at given stops during specific time intervals. Experimental results demonstrate that addressing the class imbalance significantly improves model accuracy and better represents true passenger behaviour. Additionally, Deep-GAN outperforms traditional data resampling techniques by generating synthetic data with greater diversity and realism, leading to stronger predictive performance. This work highlights the importance of data balancing techniques in improving travel demand forecasting and individual travel behaviour analysis.