Temporal Knowledge Graph Reasoning with Historical Contrastive Learning

Temporal knowledge graph, serving as an effective way to store and model dynamic relations, shows promising prospects in event forecasting. However, most temporal graph reasoning methods are highly dependent on the recurrence or periodicity of events, which brings challenges inferring future events related entities that lack historical interaction. In fact, current moment is often combined effect a small part information those unobserved underlying factors. To this end, we propose new forecasting called Contrastive Event Network (CENET), based novel training framework contrastive learning. CENET learns both non-historical dependency distinguish potential can best match given query. Simultaneously, it trains representations queries investigate whether depends more by launching The further help train binary classifier whose output boolean mask indicate search space. During inference process, employs mask-based strategy generate final results. We evaluate our proposed five benchmark graphs. results demonstrate significantly outperforms all existing metrics, achieving at least 8.3% relative improvement Hits@1 over previous state-of-the-art baselines event-based datasets.