Pseudo-Edge: Semi-Supervised Link Prediction with Graph Neural Networks

Hyunjin Seo, Seongjun Yun, Buru Chang, Jaewoo Kang

Keywords: Graph Neural Networks, Link Prediction, Pseudo-labeling, Semi-supervised learning

Abstract: Pseudo-labeling is one of the powerful Semi-Supervised Learning (SSL) approaches, which generates confident pseudo-labels of unlabeled data and leverages them for training. Recently, pseudo-labeling has been further extended to Graph Neural networks (GNNs) to address the data sparsity problem due to the nature of graph-structured data. Despite their success in the graph domain, they have been mainly designed for node-level tasks by utilizing node-level algorithms (e.g., Label Propagation) for pseudo-labeling, which can not be directly applied to the link prediction task. Besides, existing works for link prediction only use given edges as positively-labeled data, and there have been no attempts to leverage non-visible edges for training a model in a semi-supervised manner. To address these limitations, we revisit the link prediction task in a semi-supervised fashion and propose a novel pseudo-labeling framework, Pseudo-Edge, that generates qualified pseudo-labels in consideration of graph structures and harnesses them for link prediction. Specifically, our framework constructs distance-based potential edge candidates and carefully selects pseudo-labels through our relation-aware pseudo-labels generation, which reflects the comparative superiority of each unlabeled edge over its local neighborhoods in graphs. Also, we propose uncertainty-aware pseudo-labels generation that can effectively filter out over-confident samples when the model overfits to specific graph structures. Extensive experiments show that our method achieved remarkable performance across five link prediction benchmark datasets and GNN architectures, compared to state-of-the-art GNN-based semi/self-supervised models.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning