Know Your Neighbors: Subgraph Importance Sampling for Heterophilic Graph Active Learning
Keywords: Heterophilic Graph, Active Learning
Abstract: Graph neural networks (GNNs) have shown superiority in various data mining tasks but rely heavily on extensively labeled nodes. To improve the training efficiency and select the most valuable nodes as the training set, graph active learning (GAL) has gained much attention. However, previous GAL methods are designed for homophilic graphs, and their effectiveness on heterophilic graphs is less examined. In this paper, we study active learning on heterophilic graphs, where nodes with the same labels are less likely to be connected. We are surprised to find that *previous GAL methods fail to outperform the naive random sampling on heterophilic graphs*. Through an insightful investigation, we find that previous GAL-selected training sets imply homophily even on heterophilic graphs, leading to their defectiveness. To address this issue, we propose the principle of *``Know Your Neighbors''* and design an active learning algorithm KyN specifically for heterophilic graphs. The primary idea of KyN is to let GNNs receive a correct homophily distribution by labeling nodes along with their neighbors. We build KyN based on subgraph sampling with probabilities proportional to $\ell_1$ Lewis weights, which has a solid theoretical guarantee. The effectiveness of KyN is evaluated on various real-world datasets.
Primary Area: learning on graphs and other geometries & topologies
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.
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: 2517
Loading