Abstract
Abstract
Temporal Knowledge Graph Completion (TKGC) aims to model dynamic facts by incorporating the temporal dimension, so as to predict missing entities or relations under specific temporal conditions. Compared with static knowledge graph completion, TKGC requires not only learning the structural semantics of entities and relations, but also capturing the dynamic patterns of knowledge evolution over time. Existing TKGC methods generally employ fixed temporal representation mechanisms, which have limited capability in capturing complex temporal patterns and periodic evolutionary characteristics. Meanwhile, their scoring modules rely merely on shallow linear operations to fuse entity, relation and temporal features, lacking deep nonlinear interaction capabilities and thus struggling to capture complex cross-period dependencies, ultimately restricting the performance of temporal link prediction. To address these issues, we propose TeNS, a Temporal Neural Scoring model for TKGC. The core mechanism generates continuous temporal representations with a learnable time encoder. A Multilayer Perceptron (MLP) scoring network is then used to capture deep nonlinear interactions across head entity, relation and temporal features. We further optimize model training by adopting random negative sampling together with binary classification loss. Experiments on multiple public benchmark datasets demonstrate that the proposed TeNS outperforms mainstream TKGC baselines. Ablation experiments further validate the effectiveness of the learnable time encoder, MLP scoring network and random negative sampling strategy for TKGC.