A temporal knowledge graph completion method based on feature interaction enhancement

Abstract

Abstract Temporal Knowledge Graph Completion (TKGC) aims to infer missing links within Temporal Knowledge Graphs (TKGs). Due to the inherent sparsity and dynamic nature of real-world knowledge, TKGs are often incomplete. Current TKGC methods attempt to infer missing information based on existing facts. Nevertheless, these methods suffer from two primary limitations. First, they often treat temporal information as a monolithic entity, making it difficult to capture the rich hierarchical structures and dependencies inherent in time. Second, the mechanisms for modeling the interactions among entities, relations, and time are often constrained, which restrict the full capture of complex temporal dynamics. Therefore, we propose InteractExT (Time Extension of InteractE), a novel TKGC method based on feature interaction enhancement. Concretely, we first introduce a multi-granularity temporal encoder. This encoder decomposes timestamps into fine-grained components (i.e., year, month, and day) and utilizes a sequence model to capture their hierarchical dependencies, thereby mitigating the issue of temporal sparsity. Simultaneously, independent 1D convolutions are applied to extract local patterns within each component, enhancing feature distinctions and complementarity across different temporal granularities. Furthermore, we construct a multi-element interaction convolution module. By extending the conventional feature reshaping mechanism, this module facilitates comprehensive interaction among entity, relation, and temporal features, enabling a more accurate capture of the dynamic patterns of entities and relations as they evolve. Finally, extensive experiments on four benchmark datasets demonstrate that InteractExT outperforms other state-of-the-art models, validating its effectiveness in the task of TKGC. Our code is available at https://github.com/nnnu-xys/InteractExT .