SHIKI: Self-Supervised Heuristic for Improving MLPs' Knowledge by Integrating GNNs
Keywords: GNNs, Self-Supervised Learning, Deep Learning
TL;DR: Creating edges on non-graphical data and training a GNN for provable performance improvement.
Abstract: Graph Neural Networks (GNNs) are widely recognized as leading architectures for addressing classification problems involving graphical data. In this thesis, we formally define the challenge of effectively constructing edges within a dataset and training a GNN over this graph and introduce SHIKI - a novel method to tackle this task. We provide a comprehensive theoretical analysis demonstrating how graph convolutions can improve expected performance by leveraging edges. Our study focuses on the node classification problem within a non-linearly separable Gaussian mixture model, combined with a stochastic block model, and we visually demonstrate its applicability to real-world datasets. Specifically, we show that a single graph convolution in the second layer can reduce the expected loss when applying a heuristic for edge creation. We validate our findings through extensive experiments on both synthetic and real-world datasets, including those related to the entity matching problem and textual review classification. For the synthetic data, we conduct experiments based on the dataset's difficulty and various hyperparameters in our method, drawing connections between the two. Additionally, we perform an ablation study by systematically removing components of our method and testing the resulting degraded approach, which highlights the necessity of our full method. We employ several GNN architectures in the experiments, including GCN, GraphSAGE, and GAT.
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 7614
Loading