Meta-Heuristic Hyperparametric Optimization Model BO-GCN-LSTM for Spatial Interaction-Oriented Prediction Based on Deep Learning

Understanding spatial interactions is crucial for analyzing human activities across regions, yet predicting these interactions between cities on a large scale remains challenging due to the complexity of influencing factors. This study integrates a meta-heuristic optimization algorithm with a deep learning model to enhance the accuracy of such predictions nationwide. A Graph Convolutional Network (GCN) is employed to aggregate geographic features and capture spatial dependencies, while a Long Short-Term Memory Network (LSTM) identifies temporal patterns. To optimize the models, a Bayesian optimization algorithm (BO) and Genetic Algorithm (GA) were utilized to search for optimal hyperparameters, resulting in the BO-GCN-LSTM and GA-GCN-LSTM models. Comparative analysis reveals that the MSE values of the BO-GCN-LSTM and GA-GCN-LSTM models are 23% and 22.4% lower, the RMSE values are 23.9% and 15.7% lower, and the R2 values are 18% and 20% higher, respectively, than those of the GCN-LSTM model. These results demonstrate that the BO-GCN-LSTM model is particularly well-suited for large-scale predictions of spatial interactions between cities.