Weakly Supervised Graph Contrastive Learning

Pratheeksha Nair, Reihaneh Rabbany

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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.

Keywords: weak supervision, graph contrastive learning, noisy label learning, weakly supervised node classification

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: The paper introduces a novel weakly supervised graph contrastive learning method for weakly supervised node classification that leverages signals from weak labels and graph community structure.

Abstract:

Graph Contrastive Learning (GCL) has recently gained popularity owing to its ability to learn efficient node representations in a self-supervised manner. These representations are typically used to train a downstream classifier. In several real-world datasets, it is difficult to acquire sufficient clean labels for classification and instead we have weak or noisy labels available. There is little known about the robustness of the node representations learnt by the current GCL methods in the presence of weak labels.
Moreover, GCL has been successfully adapted to a supervised setting where class labels are used to contrast between pairs of nodes.
Can weak labels similarly be leveraged to learn better node embeddings? In this paper, we first empirically study the robustness of current GCL node representations to weak supervision. Then, we introduce Weakly Supervised Graph Contrastive Learning, WSNet, a novel method that incorporates signals from weak labels for the contrastive learning objective. We evaluate WSNet on five benchmark graph datasets comparing its performance with state-of-the-art GCL and noisy-label learning methods. We show that WSNet outperforms all baselines particularly in the high noise setting. We conclude that although current GCL methods show great promise in the weak supervision paradigm, they are still limited in their capacity to deal with label noise and utilizing signals from weak labels is an effective way to improve their performance.

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.

Submission Number: 6809