Multi-modal graph representations for robust anti-pattern detection in evolving codebases

Abstract

This study examines whether multi-modal and multi-level representations enhance the reliability of code smell and anti-pattern detection in evolving polyglot software systems. A hybrid model is introduced that integrates four evidence channels – structural, semantic, metric, and evolutionary – within a unified Code Property Graph (CPG) combining AST, CFG, and PDG relations. Semantic information is obtained from pretrained code language models, while classical quality indicators (e.g., CK, McCabe/Halstead) are attached as node and edge attributes; version-control signals (e.g., churn, co-change, recency) are aggregated with time decay to emphasize recent activity. Learning proceeds hierarchically: a local encoder summarizes token-level idioms and induced graph slices; a component-level, relation-aware GNN captures cohesion/coupling and data/control-flow structure; and a project-level encoder propagates context on a component-interaction graph. Instance-wise channel gating is employed to weight modalities, thereby emphasizing source-specific and smell-specific evidence. To support deployment under open-world conditions, selective prediction is adopted using complementary uncertainty criteria (logit energy, predictive entropy, stochastic variance), with temperature calibration to improve probability reliability and enable abstention on unfamiliar or low-confidence cases. The empirical evaluation spans Java, Kotlin, and Scala repositories under crossproject and time-aware splits; open-set tests are formed by withholding one smell class during training. Relative to rule/metric baselines, AST-GNN, text-only, and AST+Text systems, the hybrid model yields consistent improvements without increasing FPR@95TPR. Averaged over repositories, Macro-AUPRC improves by approximately 6–7 percentage points and Macro-F1 by 3–4 points over the strongest single-view baseline, with the largest gains observed for God Class and Shotgun-Surgery–like categories. Incremental CPG updates and bounded project-level propagation maintain CI/CD-compatible latency, while hierarchical attention and channel-importance scores provide reviewer-aligned explanations. The findings indicate that smells are inherently multi-signal and context-dependent, and that hierarchical, calibrated, open-set detection offers a favorable balance between accuracy and operational safety.