An Empirical Analysis of Constraint-Guided Embedding Dynamics for Unsupervised Anomaly Detection in Knowledge Graphs

Knowledge graph quality can be examined through statistical plausibility and structural constraints. This study investigates when embedding models and constraint-guided continuous-time dynamics provide complementary information for detecting anomalous triples in static knowledge graphs. Experiments on FB15k-237, CoDEx-Medium and WN18RR compare four knowledge graph embedding (KGE) models with deterministic and stochastic dynamics under random substitutions and constraint-oriented perturbations. KGE pipelines provide stronger recovery under random substitution; other comparisons depend on the dataset, configuration and decision rule, with dynamical scores frequently near or below chance. In WN18RR asymmetry at 15%, ten perturbation realizations confirm a mean normalized discounted cumulative gain (NDCG) improvement of 0.0249 from fusion with TransE, together with enriched exclusive recovery. Head-degree fusion performs better in the same condition, delimiting the advantage relative to simple structural alternatives. Controls using the same symbolic checker and full-graph training reveal further dataset dependence. Sequential ablations distinguish contributions from data evolution, constraint forces and symbolic checking. An exploratory expert study examines agreement in applying a diagnostic taxonomy; no generated explanations are evaluated. The findings characterize reproducible complementary recovery and its structural context, informing within-graph triage and further investigation of graph-dependent dynamical responses.