DFDM: Temporal graph functional dependency mining via spatiotemporal modeling

Temporal graph functional dependencies (TGFDs) are fundamental for enforcing data consistency, uncovering latent patterns, and enabling reliable inference in temporal knowledge graphs (TKGs). As TKGs are increasingly adopted in domains such as social networks, healthcare, and intelligent transportation, maintaining data quality is critical. Their evolving nature, however, renders them vulnerable to inconsistencies and noise, which can impair downstream reasoning, prediction, and constraint validation tasks. To address this challenge, this paper proposed dynamic temporal graph functional dependency mining (DFDM), a TGFD discovery algorithm tailored for TKGs. DFDM jointly captures structural and temporal patterns while preserving semantic consistency across evolving snapshots. It consists of four key modules: an adaptive temporal segmentation module for adjusting temporal granularity; a spatiotemporal decay graph convolution network (STD-GCN) module that captures localized multi-hop dependencies modulated by temporal proximity; a dependency-aware constrained gated recurrent unit (DC-GRU) module that models long-range temporal dynamics with constraint-aware memory; a similarity-guided dependency discovery module leveraging feature extraction. Ablation studies confirm the complementary strengths of STD-GCN and DC-GRU, highlighting the importance of jointly modeling spatial and temporal features. DFDM achieves excellent accuracy, robustness, and generalization, and it can effectively detect abnormal device behaviors.