Abstract

Knowledge graphs represent an unparalleled opportunity for machine learning, given their ability to provide meaningful context to the data through semantic representations. However, general-purpose knowledge graphs may describe entities from multiple perspectives, with some being irrelevant to the learning task. Despite the recent advances in semantic representations such as knowledge graph embeddings, existing methods are unsuited to tailoring semantic representations to a specific learning target that is not encoded in the knowledge graph

We present evoKGsim+, a framework that can evolve similarity-based semantic representations for learning relations between knowledge graph entity pairs, which are not encoded in the graph. It employs genetic programming, where the evolutionary process is guided by a fitness function that measures the quality of relation prediction. The framework combines several taxonomic and embedding similarity measures and provides several baseline evaluation approaches that emulate domain expert feature selection and optimal parameter setting.