Formulating Node Labelling as Node Classification or Link Prediction in Different Graph Representations

TMLR Paper4771 Authors

Abstract: Message-passing Graph Neural Networks (GNNs) are increasingly used for predictive tasks on graphs. Much work has been done to improve GNN architectures, but how the actual data graph should be designed is not well studied. In this paper, we investigate how two different graph representations impact the performance of GNN models across datasets with varying characteristics grouped by homophily, heterogeneity, and number of labels per node. A unique phenomenon is that the same abstract predictive task of labelling nodes is formulated as a node classification problem on one representation and as a link prediction problem on the other. Our work is the first to blur the line between these two basic and fundamental tasks in graph learning. Our experiments on 12 real-world datasets suggest that different representations (and tasks) are optimal for different datasets, models, and hyperparameters. We derive empirical heuristics of choosing between the two and pave the way towards a criterion of choosing the optimal graph representations and towards formally understanding the interconnection between node classification and link prediction.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Moshe_Eliasof1

Submission Number: 4771