FairGIN Predicts Bike-Sharing Demand Equitably for Expanding Systems

Bike-sharing systems are crucial for sustainable urban transport, but their expansion faces two key challenges: predicting demand for newly deployed stations with no historical data (cold-start problem) and ensuring equitable access, as historical data might reflect existing inequalities. Traditional models, if trained directly on such data, could inadvertently reinforce these disparities. To address this, a new framework called FairGIN has been introduced. This graph neural network is designed to be fairness-aware, integrating three main components. First, it uses Expansion-Simulated Increment Training to mimic network growth during training, bridging the gap between training and inference for evolving networks. Second, Attention-Based Knowledge Transfer allows the model to leverage data from established stations to inform predictions for new, data-sparse ones. Finally, Fairness-Aware Optimization incorporates income-stratified regularization and an equity-calibrated deployment score, promoting more inclusive station placement. Evaluations on bike-sharing data from New York City and Seattle demonstrate that FairGIN not only achieves state-of-the-art predictive accuracy across various expansion scenarios but also significantly reduces income-based disparities in access without compromising the overall efficiency of the system.