Abstract
Causal discovery from event pairs is essential for understanding complex real-world phenomena. Large language models (LLMs) have shown strong capabilities in capturing the semantics of events and inferring plausible cause-effect relations from text. However, they typically process each event pair in isolation and struggle to model the global event structure, which limits their ability to capture interdependencies among multiple events. Graph-based methods offer a structural alternative by explicitly modeling connections between events, but they often lack relational expressiveness, as relations are treated as implicit edges rather than as entities. Homogeneous hypergraphs address this by representing relations as nodes, enabling richer modeling of multi-event interactions and more expressive causal reasoning. Nevertheless, this strategy frequently leads to disconnected structures, hindering information aggregation through graph neural networks (GNNs). To address these challenges, we propose eCHOLGA (edge Classification through Heterogeneous One-cLass Graph Autoencoder), a novel method that leverages heterogeneous hypergraphs to model causal relationships more effectively. eCHOLGA integrates semantic features extracted from language models into the graph structure, enhancing the representation of events and their relations. By transforming relations into nodes and introducing additional node and edge types, it improves topological connectivity and enables GNNs to learn more informative edge representations. Furthermore, our method adopts a one-class learning strategy, requiring only positive (causal) examples for training, which reduces labeling effort. In addition to its effectiveness, eCHOLGA enhances interpretability and provides insights into the causal discovery process. Experimental results show that eCHOLGA outperforms state-of-the-art methods, establishing it as a promising approach for causal discovery in event pairs.