Causal Graph Representation Learning for Outcome-Oriented Link Prediction
Langjunqing Jin, Feng Zhao, Cheng Yan, Xiangyu Gui · 2024
Causal graph representation learning has been an essential tool for improving the performance achieved in downstream tasks such as link prediction in causal graphs. Most of the existing methods can learn the links between binary nodes in an observed graph structure. However, the causal relations between nodes and their multivariate continuous nature are easily overlooked when learning knowledge graph links. This can lead to difficulty when attempting to find the real causes affecting the facts in multiple node chains, thus reducing the accuracy of prediction. Therefore, we propose a causal representation learning model for link prediction (CRLLP) to accurately complete missing links in causal graphs. Treating the node embeddings in causal graphs as contexts is a novel paradigm for characterizing causal relations in which the exsiting graph structure are treatments and the links to be completed are outcomes. In addition, we propose an outcome-oriented counterfactual link prediction method that learns causal chain representations based on observed nodes and queries the causal links in causal graphs by answering counterfactual questions. The developed method can clearly identify the causes corresponding to different outcomes. The experimental results show that our model can achieve optimal link prediction and causal chain inference performance with highly accurate causal characterization representations. Furthermore, other experiments demonstrate the superiority and interpretability of our model.