Multiplex graph prompt learning and attentive fusion for event graph completion

This paper introduces an Event Graph Completion (EGC) task to predict the absence of multi-relations between events in a heterogeneous Event Graph (EG) with four prevailing relations. The primary objective of the EGC task is to enhance the completeness of the heterogeneous EGs, thereby optimizing their structure and improving their utility for downstream applications. Considering that multi-relations may facilitate the completion of each other, that is, completing one type of event relations may benefit from the other type of relations. We propose a Multiplex Graph Prompt Learning and Attentive Fusion (PLAF) model to learn event representations interactively among each homogeneous graphs and adopt a kind of dual graph prompt learning for missing relation prediction. The PLAF model includes three key modules: (1) Dual Graph Prompt Learning (DGPL) reformulates each EG into an event triplet sequence to encode its structural and semantic information; (2) Multiplex Graph Attention Network (MGAT) divides each heterogeneous EG into four homogeneous graphs to learn events' representations through both inter-graph and cross-graph attention; (3) Aggregative Relation Prediction Module (ARPM) aggregates the missing relation predictions from both DGPL module and MGAT module to complete the EGs. Furthermore, we have constructed the EGC-MAVEN dataset and conducted extensive experiments to evaluate the efficacy of our approach. The experimental results validate our arguments and demonstrate that our proposed PLAF model outperforms the advanced competitors.