Mining differential causal rules in knowledge graphs

Overview & Synopsis

The mining of association rules within knowledge graphs is an important area of research.Indeed, this type of rule makes it possible to represent knowledge, and their application makes it possible to complete a knowledge graph by adding missing triples or to remove erroneous triples.However, these rules express associations and do not allow the expression of causal relations, whose semantics differ from an association or a correlation.In a system, a causal link between variable A and variable B is a relationship oriented from A to B. It indicates that a change in A causes a change in B, with the other variables in the system maintaining the same values.Several frameworks exist for determining causal relationships, including the potential outcome framework, which involves matching similar instances with different values on a variable named treatment to study the effect of that treatment on another variable named the outcome.In this thesis, we propose several approaches to define rules representing a causal effect of a treatment on an outcome.This effect can be local, i.e., valid for a subset of instances of a knowledge graph defined by a graph pattern, or average, i.e., valid on average for the whole set of graph instances.The discovery of these rules is based on the framework of studying potential outcomes by matching similar instances and comparing their RDF descriptions or their learned vectorial representations through graph embedding models.