<![CDATA[Software modularization remains an issue until this day. As software grows over time, software becomes more complex and difficult to maintain. To address this issue, software module clustering was used by grouping software components such as classes, objects and files. Many approaches have been conducted by using single-feature clustering and multi-features clustering. The result shows that multi-features clustering performs better than single-feature clustering. The application of GNN also gains significant attention for software modularization and proven to produce better performance. In this study, we analyze how integrating structural features, semantic features and graph relations using Attention-drive Graph Clustering Network (AGCN) affects the quality of software clustering. We use four input combinations which are structural features with graph, semantic features with graph, identity matrix with graph, and the fusion of graph, semantic and structural features across all three java projects (JavaCC, Dom4J, and JEdit). The result shows that by using the structural feature only the model could predict the total cluster that is closest to the original total cluster and multi-feature fusion achieves the best NMI score across all input variants.]]>
Copyrights © 2026
