Graph Pruning Based Spatial and Temporal Graph Convolutional Network with Transfer Learning for Traffic Prediction

TL;DR

TL-GPSTGN tackles traffic prediction under data-scarce conditions by integrating graph pruning with transfer learning in a spatial-temporal graph convolutional framework. The Graph Pruning Processor distills the road network via information entropy and degree-based pruning, enabling robust migration of a pre-trained STGCN to new networks with limited data. Empirical results on METR-LA, PEMS-BAY, and PEMS-D7-M show competitive single-dataset accuracy and superior transfer performance, highlighting strong generalization and deployment potential. The theoretical analysis via Rademacher complexity supports the model’s generalization capacity, and the framework offers practical utility for scalable, transferable smart transportation systems.

Abstract

With the process of urbanization and the rapid growth of population, the issue of traffic congestion has become an increasingly critical concern. Intelligent transportation systems heavily rely on real-time and precise prediction algorithms to address this problem. While Recurrent Neural Network (RNN) and Graph Convolutional Network (GCN) methods in deep learning have demonstrated high accuracy in predicting road conditions when sufficient data is available, forecasting in road networks with limited data remains a challenging task. This study proposed a novel Spatial-temporal Convolutional Network (TL-GPSTGN) based on graph pruning and transfer learning framework to tackle this issue. Firstly, the essential structure and information of the graph are extracted by analyzing the correlation and information entropy of the road network structure and feature data. By utilizing graph pruning techniques, the adjacency matrix of the graph and the input feature data are processed, resulting in a significant improvement in the model's migration performance. Subsequently, the well-characterized data are inputted into the spatial-temporal graph convolutional network to capture the spatial-temporal relationships and make predictions regarding the road conditions. Furthermore, this study conducts comprehensive testing and validation of the TL-GPSTGN method on real datasets, comparing its prediction performance against other commonly used models under identical conditions. The results demonstrate the exceptional predictive accuracy of TL-GPSTGN on a single dataset, as well as its robust migration performance across different datasets.

Figures (6)

• Figure 1: Model Structure of TL-GPSTGN. The framework utilizes a transfer learning approach with spatio-temporal graph convolutional networks (STGCN) for traffic prediction. The architecture processes source and target graph data through Graph Preprocessing Pipelines (GPP) and spatio-temporal convolutional layers (ST-Conv) with parameter sharing. The model outputs predictions ($X_{pred}$) using feature reductor layers, supported by input preprocessing (IEA, GP, Norm) and task-specific layers (TCL, GCL, Linear).
• Figure 2: Graph Pruning Visualization. This figure shows the raw and processed adjacency matrices for the metr-la (top) and pemsd7-m (bottom) datasets. Graph pruning reduces noise and highlights significant connections, improving sparsity and enhancing the model's ability to capture meaningful spatial relationships.
• Figure 3: Loss of TL-GPSTGN trained on metr-la dataset. Train and test loss for 15-minute (left) and 30-minute (right) prediction intervals. Both plots show rapid loss reduction in early epochs, followed by stabilization, highlighting the model's convergence and generalization ability.
• Figure 4: Transfer learning loss on pemsd7-m dataset. Loss curves for transfer learning on 15-minute (left) and 30-minute (right) intervals. Both plots demonstrate the model's effective adaptation to the target dataset, with steady loss reduction and convergence, showcasing successful knowledge transfer.
• Figure 5: Partial prediction results on the pemsd7-m datasets. Predictions for 15-minute (left) and 30-minute (right) intervals on the Pemsd7-m dataset using transfer learning. The GPSTG model aligns closely with the ground truth, outperforming the baseline STGCN model in capturing traffic patterns.