Deep Semantic Embeddings for Scalable Co-Change Recommendation in Software Systems

Accurate co-change prediction is essential for preventing regressions and estimating maintenance costs. Most software co-change prediction models excel at identifying obvious dependencies but fail to capture complex, multi-file relationships, leading to poor performance at larger recommendation sizes. We propose a solution that shifts the focus from historical patterns to semantic code compatibility using pre-trained transformers (GraphCodeBERT) and supervised learning. Validated against three Java repositories, our approach demonstrates a fundamental methodological advantage: it achieves monotonic performance improvement as the recommendation window expands. By effectively identifying the "long-tail" of hidden dependencies, this semantic recommendation framework provides a more reliable tool for change impact analysis and resource allocation in large-scale software evolution.