Temporal knowledge graph completion via local–global contrastive embedding learning

Temporal knowledge graphs (TKGs) consist of timestamped facts and serve as a core substrate for many AI applications; their completeness directly influences downstream performance. However, real-world TKGs are typically incomplete, making the inference of missing facts from observed ones a central research problem. Current temporal knowledge graph completion (TKGC) methods predominantly rely on embedding-based models that map entities, relations, and timestamps into low-dimensional vector spaces to predict missing facts. Although attention mechanisms have been employed to encode structural information, most approaches focus on local neighborhoods around entities or relations, thereby overlooking the global graph structure and under-exploiting temporal patterns at multiple time scales. To address these limitations, a new TKGC model based on local-global contrastive embedding is introduced. At the global level, self-attention is used to encode entities and relations to capture long-range structural dependencies across the graph. At the local level, positional encodings and temporal-duration embeddings are integrated with self-attention layers to mine fine-grained temporal signals carried by timestamps, enabling a comprehensive characterization of how entities and relations evolve over time. In addition, inspired by contrastive learning, global structural representations are contrasted and fused with local temporal representations, which effectively align and strengthen entity and relation embeddings. Evaluations on three widely used benchmarks—ICEWS14, ICEWS05-15, and GDELT—demonstrate consistently strong performance across metrics, confirming the robustness of the proposed approach for TKGC.