Discovering causal relationships between different observations is the goal of many experiments in science. When observational data are available, the potential outcome framework is a well-used framework for discovering such relationships. To study the effect of a treatment T, similar individuals can be examined to estimate counterfactuals. In this work, we adapted this framework to be used in Knowledge Graphs (KGs) to discover causal rules that express that differences in treatments lead to differences in a studied characteristic.
However, quantifying the similarity between individuals represented in a knowledge graph is challenging, especially because their descriptions can be incomplete and erroneous. We develop a new matching method based on knowledge graph embeddings. We have also develop a new algorithm that uses this new matching method to discover causal rules in KGs. The experiments that we conducted on a scientific knowledge graph, by involving a domain expert, showed that our approach is able to discover rules that explain much more differences in the studied characteristic, than existing state of the art approaches.