Zero-Shot Knowledge Graph Completion for Recommendation System

Abstract

Knowledge graphs are structured representations of actual entities and relations. They are widely used to improve the performance of downstream tasks such as recommendation systems and semantic searching. Knowledge graph completion (KGC) is a technology for discovering the missing relations between the entities in a knowledge graph (KG). Existing methods leverage known relations on a KG to build a model to predict missing relations. Such methods implicitly require a substantial number of relations to be known in advance, which might not be available in practice. To cope with the cold-start scenario for KGC, i.e., no relation is known in advance, we propose a zero-shot approach in this paper. Our approach converts the KGC process to an optimization problem. It uses the Evolutionary Strategy (ES) algorithm to optimize a model used to complete the KG according to the performance of the recommendation system constructed based on the completed KG. Experiments on a movie dataset demonstrate that our approach can complete the KG in the cold-start scenario and improve the performance of the recommendation system built based on the completed KG.