One-class edge classification through heterogeneous hypergraph for causal discovery.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Graph Neural Networks

Background:

• Causal discovery from event pairs is crucial for understanding complex systems.
• Large language models (LLMs) excel at event semantics but struggle with global event structure.
• Existing graph-based methods lack relational expressiveness and can lead to disconnected structures.

Purpose of the Study:

• To propose a novel method, eCHOLGA (edge Classification through Heterogeneous One-cLass Graph Autoencoder), for effective causal discovery from event pairs.
• To leverage heterogeneous hypergraphs and LLM semantic features for enhanced causal relationship modeling.
• To improve topological connectivity and enable informative edge representations using graph neural networks (GNNs).

Main Methods:

• eCHOLGA utilizes heterogeneous hypergraphs, transforming relations into nodes and incorporating additional node/edge types.
• Semantic features from LLMs are integrated into the graph structure for richer event and relation representation.
• A one-class learning strategy is employed, requiring only positive causal examples for training, reducing labeling effort.

Main Results:

• eCHOLGA demonstrates superior performance compared to state-of-the-art methods in causal discovery tasks.
• The method enhances topological connectivity, enabling GNNs to learn more informative edge representations.
• Experimental results validate the effectiveness of integrating LLM features and heterogeneous hypergraphs.

Conclusions:

• eCHOLGA offers a promising approach for causal discovery in event pairs by effectively modeling complex interdependencies.
• The method improves causal reasoning, interpretability, and reduces the need for extensive labeled data.
• This work advances the field by providing a more expressive and connected graph-based framework for causal inference.