Gcn-Gat-Mlp: A Hybrid Geographic-Semantic Context-Aware Urban Commuting Flow Prediction Model Using Neural Networks

Prediction of urban-scale commuting flow plays an important role in urban and transportation planning and optimization. The traditional and machine-learning solutions used for this purpose have some limitations. Traditional models, such as gravity, are analytical physics-based models that lack important origin and destination features like land use and sociodemographic information. On the other hand, machine learning-based models leverage the essential information of origin and destination zones, which makes them better than traditional ones; however, they mostly ignore the influence of geographically and semantically nearby zones. To address these issues, we propose a hybrid model that captures the information of the neighbor zones both geographically and semantically and improves the prediction accuracy. In the first step, we extract the features of all geographic units, such as Points of Interest (POIs) and distance to the Central Business District (CBD). Then, we use the geographic adjacency matrix to encapsulate the information of nearby zones into each zone using Graph Convolution Network (GCN). In the third step, we consider spatial flow interaction to construct the semantic adjacency matrix and use it to encode the information of semantically nearby zones using Graph Attention Network (GAT). Finally, a Multilayer Perceptron (MLP) regressor is trained to predict the commuting flow based on the embeddings and flow. We evaluate our proposed model on a real-world Montreal, CA dataset, demonstrating our model's effectiveness relative to the state of the art.