Toward accurate learning of graph representations

Abstract

While machine learning algorithms and models for graph-structured data have been actively studied, the problems of handling new entities (nodes) and representing entire graphs remain challenging. My thesis focuses on how to tackle such problematic issues on a graph. In other words, we develop methods for representing unseen entities and entire graphs more accurately, summarized into two folds: For the problem of representing unseen entities, we first introduce a realistic task of out-of-graph link prediction that aims to predict missing links for unseen entities. Then, to tackle this, we propose a transductive meta-learning framework that makes it possible to simulate the unseen during training. We validate our method on benchmark datasets for knowledge graph completion and drug-drug interaction prediction. The experimental results show that our method significantly outperforms existing baselines on the out-of-graph link prediction task, due to its effectiveness in accurately representing unseen entities. For the problem of representing entire graphs, we aim to embed different graphs into distinct vectors. To do so, we consider the graph encoding problem as a multiset encoding problem, which allows for possibly repeating elements, since a graph may have redundant nodes. Then, over the multiset encoding scheme, we propose a graph multiset transformer that captures interaction among nodes, while reducing the size of the given graph, to obtain a compact yet entire graph representation. We theoretically prove that our method is as powerful as the Weisfeiler-Lehman graph isomorphism test, but also empirically show that it outperforms baselines on graph classification, reconstruction, and generation tasks. We believe both of our approaches contribute to the optimal goal of accurate learning of real-world graphs, often evolving with unseen nodes and having a large number of nodes to capture at once.