CompNet : A novel Knowledge Graph Embedding Technique for Link Prediction

Abstract

A Knowledge Graph(KG) is a graph that contains information about anything and everything in the world, these graphs can be represented by plain-simple nodes and links. Knowledge Graph Embedding refers to attaining valuable information about every node present in the graph, essentially it could be defined as representing a node as a low- dimensional vector. Knowledge graph Embedding techniques have become an increasingly popular research topics. Despite all the efforts invested in creation and maintenance of knowledge graphs only contain a part of what it contains is true and it also still consists of missing facts. Prediction of these missing facts in scientific term can be known as Link Prediction. Several recent works use deep learning approaches to generate richer and more expressive embedding. However, observations show that the following methods are very expensive computationally. This paper proposes a novel approach for generating Graph Embeddings which can be used to perform the task of Link Prediction i.e. interpreting missing data. ’CompNet’ uses a modified version of the complEx and Cluster-Graph Convolution Network (Cluster-GCN) algorithms. Laying out certain constraints on the ComplEx algorithm i.e. discussed in detail through this paper, can improve the time complexity drastically. ’CompNet’ is tested on the large-scale dataset Freebase, wherein it out-performs the traditional approaches (translation based approaches and semantic search approaches) in terms of Mean Reciprocal Rank (MRR) and also utilizes low storage space.