Machine learning based semantic link prediction for biomedical knowledge graphs

The construction of large knowledge bases which contain large volumes of data about real world objects and their relationships has been an object of great interest in recent times. Nowadays, many databases in the biomedical field are accessible through Semantic Web technologies, so these databases have been presented as knowledge graphs in the form of Linked Data. The knowledge graphs are powerful models for defining data which also enable the application of graph mining algorithms thanks to the underlying graph structure. A number of approaches for the discovery of missing links and prediction of new links in large knowledge graphs is presented within this thesis. A hybrid approach based on machine learning which uses knowledge graphs to discover new links between entities in biological and biomedical information network is proposed. Two groups of features, local and global features, based on structural and semantic properties of the knowledge graphs, are used for the link prediction. Local features are based on network proximity and global features are based on vector representation of the semantic graph. The machine learning models trained with these two feature groups were evaluated separately and jointly. In addition, test scenarios were developed for cases that were often ignored in evaluating link prediction methods, and the proposed methods applied to these scenarios were tested. The utility of the proposed approaches was demonstrated by successfully applying them on two important problems for the drug discovery process in the biomedical field; predicting new drug-drug interactions and predicting novel drug indications. The method's superiority over existing approaches were thereby demonstrated.