Consensus Fusion of Structural and Behavioural Signals for Software Co-Change Prediction

Predicting which files must co-change in a software commit is critical for reducing incomplete pull requests and integration failures. We introduce a consensus fusion approach that combines Katz centrality on import graphs (structural signal) with co-change frequency from version history (behavioral signal) using CombMNZ, a rank fusion method from information retrieval. Evaluated on 15 open-source Python repositories (688 commits), CombMNZ achieves +57% F1 (p=0.0001) over structural analysis alone and +16% F1 (p=0.012) over the ROSE association rule mining baseline. We prove that ROSE confidence-based ranking is mathematically equivalent to raw co-change frequency under single-seed evaluation, and show that 70.9% of correct predictions come exclusively from structural analysis while 21.6% come from history alone — confirming true signal complementarity. Our work transfers the chorus effect from information retrieval to software engineering.