INTEGRATE: Distance based Graph Convolutional Networks for Statistical Relational Learning
Abstract: Recently, several successful methods for learning embeddings of large knowledge bases have been developed. They have been motivated through the inevitability of learning and reasoning about various entities, their attributes and relations present in the knowledge bases. A potential limitation of much of this line of research is that the inherent semantic structure of the network is not exploited. To overcome this limitation, graph convolutional networks (GCNs) were proposed that generalized neural network models to multi-relational, graph-structured data sets. We consider the problem of learning distance-based Graph Convolutional Networks (GCNs) for multi-relational data within statistical relational learning. Specifically, we first embed the original graph into the Euclidean space R^m using a relational density estimation technique thereby constructing a secondary Euclidean graph. The graph vertices correspond to the target triples and edges denote the Euclidean distances between the target triples. We emphasize the importance of learning the secondary Euclidean graph and the advantages of employing a distance matrix over the typically used adjacency matrix. Our comprehensive empirical evaluation demonstrates the superiority of our approach over 15 approaches spread over different GCN models, relational embedding techniques, rule learning techniques and relational models.
Submission Length: Long submission (more than 12 pages of main content)
Assigned Action Editor: ~Rémi_Flamary1
Submission Number: 985
Loading