| Summary: | Although machine learning has shown promising results and attracted great attentionin the recent years, few methods can be applied to relational, or graphstructured,data. Knowledge Graph Embedding (KGE) is a learning methodology that attemptsto bridge this gap by learning to represent nodes as vectors that can be used for variousdownstream tasks. A multirelational, or knowledge, graph, i.e. a graph with multipletypesofedges,isageneralabstractionthatcanrepresentpairwiserelationshipsbetweena set of entities such as friendships in a social network, citations and coauthorshipin an academic network or even various types of atomic bonds in a molecule. In therecent years, a number of large knowledge bases have been assembled that containup to several billions of entities and facts about them, represented as links betweenthem. Since most large knowledge graphs are inherently incomplete, link prediction,which is the identification of missing links, is the standard task on which KGE modelsare evaluated and where they are able to achieve stateoftheart performance. Entityclassification, where the goal is to predict the class an entity belongs to, is anotherpotential application of KGE models which has been less explored. For this thesis IimplementedaKGElibrarywhichIusedtoevaluatevariousmodelsandmethodologieson link prediction for the benchmark datasets FB15K and WN18 as well as entityclassification for spammer detection in a social network.
|
|---|
