Multiplex graph prompt learning and attentive fusion for event graph completion.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Heterogeneous Event Graphs (EGs) often suffer from missing relational information, limiting their downstream application utility.
• Existing methods struggle to effectively model and predict multiple, interconnected relations within EGs.

Purpose of the Study:

• To introduce and address the novel Event Graph Completion (EGC) task for predicting absent multi-relations in heterogeneous EGs.
• To propose a new model, Multiplex Graph Prompt Learning and Attentive Fusion (PLAF), for enhancing EG completeness.

Main Methods:

• Developed the PLAF model, incorporating Dual Graph Prompt Learning (DGPL) and a Multiplex Graph Attention Network (MGAT).
• DGPL encodes EG structure and semantics via event triplet sequences.
• MGAT learns event representations using inter-graph and cross-graph attention across homogeneous subgraphs.
• An Aggregative Relation Prediction Module (ARPM) combines predictions for robust completion.

Main Results:

• The PLAF model demonstrated superior performance in predicting missing relations compared to state-of-the-art methods.
• Extensive experiments on the newly constructed EGC-MAVEN dataset validated the model's efficacy.
• The results confirm that modeling multi-relations interactively improves prediction accuracy.

Conclusions:

• The proposed PLAF model effectively addresses the Event Graph Completion task.
• The approach enhances the completeness and utility of heterogeneous event graphs for various applications.
• This work establishes a new benchmark for multi-relation prediction in event graphs.