Temporal knowledge graph (TKG) reasoning has emerged as a pivotal approach in event prediction. An important yet challenging task in TKG reasoning is to predict future events by extrapolating from historical events and their correlations. Existing methods either overlook the modeling of long-term dependencies between entities or are ineffective in aggregating long-term information with recent facts. Motivated by dual process theory in cognitive sciences, we introduce TKG-LDG, an approach enhancing TKG for future entity prediction with long-term dense graph, to model event evolution in an adaptive manner. We first construct a unified dense graph from historical data to capture long-term dependencies, reflecting cumulative knowledge of entity interactions over time. This unified dense graph is compatible with any graph neural network and facilitates entity interaction learning from a long-term perspective. Then we initialize a TKG encoder from the unified dense graph to enhance short-term event interaction modeling. TKG-LDG effectively marries global context with local adaptability to recent temporal changes through its short-term recurrent encoders, in a way that mirrors human reasoning by integrating both long-term and short-term event dynamics. Extensive experiments conducted on six widely used TKG datasets demonstrate that our model outperforms strong baselines in future event prediction.

The ability to forecast future events brings great benefits for society and cyberspace in many public safety domains, such as civil unrest, pandemics and crimes. The occurrences of new events are often correlated or dependent on historical and concurrent events. Many existing studies learn event-occurring processes with sequential and structural models, which, however, suffer from inefficient and inaccurate prediction problems. To better understand the event forecasting task and characterize the occurrence of new events, we exploit the human cognitive theory from the cognitive neuroscience discipline to find available cues for algorithm design and event prediction. Motivated by the dual process theory, we propose a two-stage learning scheme for event knowledge mining and prediction. First, we screen out event candidates based on historical inherent knowledge. Then we re-rank event candidates by probing into the newest relative events. Our proposed model mimics a sociological phenomenon called “the chameleon effect” and consists of a new target attentive graph collaborative learning mechanism to ensure a better understanding of sophisticated evolution patterns associated with events. In addition, self-supervised contrastive learning is employed to alleviate the over-smoothing problem that existed in graph learning while improving the model’s interpretability. Experiments show the effectiveness of our approach.