Footfall Prediction Using Graph Neural Networks

Accurately predicting the potential foot traffic for a new business is a crucial task since it directly impacts a business's ability to generate revenue. In this work, a graph neural network-based approach is introduced in which the foot traffic between businesses and neighborhoods is represented in a bipartite network setting where edges capture the yearly-aggregated foot traffic quartile labels. Resulting bipartite networks are fed to the graph neural network to predict the foot traffic label for a new business for all the available neighborhoods. The graph neural network model outperforms well-established Huff model by 3% higher F1 score. Our results indicate that utilizing graph neural network architectures for foot traffic prediction is promising and requires more attention from the field.