Evolving Meaning for Supervised Learning in Complex Biomedical Domains Using Knowledge Graphs

Abstract

Knowledge graphs represent an unparalleled opportunity for machine learning, given their ability to provide meaningful context to data through semantic representations. Knowledge graphs provide multiple perspectives over an entity, describing it using different properties or multiple portions of the graph. State-of-the-art semantic representations are static and take into consideration all semantic aspects, ignoring that some may be irrelevant to the downstream learning task. The goal of this Ph.D. project is to discover suitable semantic representations of knowledge graph entities that are adapted to specific supervised learning tasks. I will use Genetic Programming to evolve tailored semantic representations, and develop novel approaches that integrate them with different supervised learning techniques. These novel approaches will be anchored by a framework that integrates different semantic representation approaches and two representative learning approaches, Support Vector Machine and Graph Convolutional Neural Networks, and allows a comparative evaluation using benchmarks. The developed approaches will be applied to two bioinformatics tasks, prediction of protein interactions and gene-disease associations, where the impact of data size and complexity will be investigated.