Abstract: Temporal knowledge graph (TKG), as an important form of knowledge representation for describing dynamic event evolution, incorporate temporal information to characterize the dynamic features of entities, relations, and facts over time, thereby providing a knowledge foundation for knowledge reasoning, event evolution analysis, and potential association discovery. To address the insufficient modeling of entity temporal evolution, inadequate utilization of historical relational dependencies, and limited dynamic semantic representation in existing temporal knowledge graph completion methods, this paper proposes a temporal knowledge graph completion method integrating temporal evolution and relational context. The proposed method consists of a Time-Conditioned Entity Encoder (TCEE), a Relation Context Aggregator (RCA), a Temporal Scoring Prediction module (TSP), and a Contrastive Learning Module (CLM). TCEE introduces time-conditioned parameters into entity representations and characterizes the dynamic features of entity representations over time through time-conditioned transformations and historical behavior decay, enabling entity representations to capture behavioral evolution patterns at different temporal stages. RCA uses the current query relation as guidance to retrieve historical interaction events within a historical time window and selectively aggregates historical contexts according to relational relevance, thereby reducing noise introduced by low-relevance historical information and enhancing the utilization of historical relational dependencies and temporal contextual information. TSP jointly models time-enhanced entity representations and time-aware relation representations, introduces an asymmetric interaction mechanism to characterize the directional relationships between subjects and objects, and performs missing fact prediction through a temporal scoring function. CLM constructs temporally consistent positive and negative samples based on entity behavioral relations at different temporal stages and constrains the representation space through contrastive learning, making entity representations with similar semantic and temporal features closer while separating those with substantially different features, thereby enhancing the discriminability and robustness of temporal knowledge representations. Experiments were conducted using publicly available event datasets ICEWS14, ICEWS05-15, and GDELT which are commonly used to model temporal interactions among entities in temporal knowledge graph completion tasks. The results show that compared with current state-of-the-art baseline models, the proposed method achieves an average improvement of 3.40% in MRR, 5.77% in Hits@1, and 1.37% in Hits@3 across the three datasets, demonstrating its effectiveness in temporal knowledge graph completion.

摘要： 时序知识图谱作为描述动态事件演化过程的重要知识表示形式，通过引入时间信息刻画实体、关系事实随时间变化的动态特征，为知识推理、事件演化分析和潜在关联发现提供基础。针对现有时序知识图谱补全方法在实体时序演化建模不足、历史关系利用不充分以及动态语义表达能力有限的基础上，提出一种融合时间演化和关系上下文的时序知识图谱补全方法，该方法由时间条件实体编码器、关系上下文聚合器、时序评分预测模块和时间一致性对比学习模块组成。时间条件实体编码器在实体表示的基础上引入时间条件参数，通过时间条件变换和历史行为衰减刻画实体表示随时间变化的动态特征，实体表示能够反映不同时间阶段的行为演化规律；关系上下文聚合器以当前查询关系为引导，在历史时间窗口内检索实体的历史交互事件，利用关系相关性对历史上下文进行选择性聚合，减少了低相关历史信息引入的噪声，增强了模型对历史关系依赖和时序上下文信息的利用能力；时序评分预测模块将时间增强实体表示与时间感知关系表示进行联合建模，引入非对称交互机制刻画主体与客体之间的方向性关系，通过时序评分函数实现缺失事实预测；时间一致性对比学习模块根据实体在不同时间阶段的行为关系构造具有时间一致性的正负样本，通过对比学习约束表示空间，使语义和时序特征相近的实体表示更加接近、差异较大的表示分离，增强了时序知识表示的判别性与鲁棒性。实验结果表明，在时序知识图谱补全任务中常用于刻画实体时序交互行为的公开事件数据集ICEWS14、ICEWS05-15和GDELT上，与当前先进基线模型相比，所提方法MRR平均提升3.40%，Hits@1平均提升5.77%，Hits@3平均提升1.37%，验证了该方法在时序知识图谱补全任务中的有效性。