CAGE-TAP: Concept-Augmented Graph Evidence for Smart-Home Trigger-Action Conflict Diagnosis

Smart-home trigger-action programming (TAP) allows users to compose automation rules, yet rule interactions can produce device clashes, environmental conflicts, trigger chains, and safety or privacy violations. Existing detectors either require manually maintained device semantics or produce opaque decisions. This paper introduces CAGE-TAP, a concept-augmented graph-evidence framework that routes every conflict prediction through a compact vector of human-checkable TAP concepts—structural concepts computed from normalized rule fields and semantic/risk concepts learned from pair-centered interaction graphs—so that explanations are faithful to the implemented decision path and correctable at inference time. We evaluate CAGE-TAP on 21,447 labelled rule pairs obtained by aligning an audited 580-pair collection TAPConflict-580 (TC-580), with IoTCOM and TAPFixer interactions. The benchmark contains 10,072 conflicts and 11,375 non-conflicts under group-disjoint train, validation, and test partitions. CAGE-TAP with an R-GCN encoder achieves 0.9280 ± 0.0004 binary F1, 0.9519 ± 0.0004 AUPRC, and 0.9020 ± 0.0035 seven-type Macro-F1. Its type diagnosis exceeds a capacity-matched direct R-GCN (0.7863 ± 0.0011) and fine-tuned BERT (0.8030±0.0049). The six learned semantic/risk concepts reach 0.8532±0.0078 Macro-F1. Targeted contribution masking changes twice as many predictions as matched random masking, while reference correction improves both binary and typed performance. Each positive conflict type is represented by at least 72 evaluated test instances; per-type F1 ranges from 0.8221 to 0.9467.