From Reactive to Proactive: Graph Convolutional Networks for Future Software Module Coupling Prediction

Software module dependency prediction is a critical task in modern software engineering for preventing future connectivity issues and improving system sustainability. This study proposes a Graph Convolutional Network (GCN) based framework to predict potential inter-module dependencies using comprehensive software metrics. Experiments were conducted on the complete NASA JM1 dataset (10,885 modules), selected for its scale and extensive use in software engineering research. All 21 software metrics were utilized without dimensionality reduction. A K-Nearest Neighbors (KNN) graph modeling approach (k=8) with a cosine similarity threshold of 0.2 captured structural relationships, producing 85,002 training edges across 15 connected components. The proposed three-layer residual GCN architecture (21→128→128→64) integrates ReLU activation, 30% dropout, and residual skip connections, along with a link-prediction-oriented data partitioning strategy. The model achieved strong performance with 97.58% AUC, 92.12% F1-score, and 99.99% recall. In addition to predictive performance, the framework demonstrated high computational efficiency, requiring an average of 0.165 seconds per training epoch and completing training in 33.1 seconds. These results indicate that the model is suitable for scalable deployment and real-time DevOps integration. By enabling proactive dependency forecasting, the proposed approach supports early identification of design risks and improves software quality management in large-scale development environments.