New graph-propagation method fuses semantic and structural impact analysis
Training-free analyzer catches blind spots in software change impact detection
Momil Seedat's new arXiv paper (arXiv:2606.18855) argues that current change-impact-analysis tools rely on either semantic similarity (via embeddings, code search) or structural dependencies (call graphs, test-impact selection) – each with characteristic blind spots. Semantic tools miss artifacts that share no vocabulary with the change; structural tools miss meaning-related artifacts without explicit edges. Most tools also operate only on code, ignoring the Requirement-Config-Service-Test chain.
Seedat's solution models the system as a heterogeneous artifact graph with typed edges from static analysis. It computes a semantic prior via cosine similarity to the changed artifact, then propagates impact multi-hop with decay using a row-normalized propagation matrix. The two signals are blended with a single tunable weight lambda. A proof-of-concept on a payment subsystem (5 labeled change scenarios) demonstrates that fusion correctly recovers both blind spots: artifacts with zero textual overlap (via propagation) and helper functions unreachable by propagation alone (via semantic layer). Lambda acts as an explicit precision/recall control. The paper also argues the method extends to operational artifacts (images, metrics, dashboards, data schemas) using four publicly documented production failures as evidence.
- Fuses semantic similarity (cosine on embeddings) with graph-propagation (typed edges, multi-hop decay) for understanding software change impact
- Proof-of-concept on a payment subsystem (5 scenarios) recovers artifacts with zero textual overlap and propagation-blind helper functions
- Training-free, interpretable, and tunable via a single lambda parameter for precision/recall control
Why It Matters
A practical, fusion-based approach that could make change impact analysis more robust across the full software lifecycle.