Graph Fairness Learning under Distribution Shifts
Keywords: Graph Neural Networks, Fairness, Distribution Shifts
Abstract: Graph neural networks (GNNs) have achieved remarkable performance on graph-structured data. However, GNNs may inherit prejudice from the training data and make discriminatory predictions based on sensitive attributes, such as gender and skin color. Recently, there has been an increasing interest in ensuring fairness on GNNs, while all of them are under the assumption that the training and testing data are under the same distribution. Will graph fairness issue also emerge under distribution shifts? How does distribution shifts affect graph fairness learning? All these open questions are largely unexplored from a theoretical perspective. In this work, we theoretically prove that graph fairness learning is determined by two key factors: the feature difference among certain groups and a fairness-related structure property of the graph. We further establish the relationship between fairness on the testing graph and two factors: fairness on the training graph, as well as the distribution difference between the training graph and the testing graph. Motivated by our theoretical analysis, we propose our framework FatraGNN to ensure fairness performance under distribution shifts on graphs. Specifically, we use a graph generator to generate graphs that result in prediction unfairness and are under different distributions. Then we maximize the alignment of representations between the training graph and generated graphs for each certain group. This empowers us to attain high classification and fairness performance even on generated graphs exhibiting significant unfairness, thereby enhancing the ability to handle the testing graphs effectively. Experiments on real-world and semi-synthetic datasets validate that our model can achieve better classification and fairness performance compared with other state-of-the-art baselines.
Track: Graph Algorithms and Learning for the Web
Submission Guidelines Scope: Yes
Submission Guidelines Blind: Yes
Submission Guidelines Format: Yes
Submission Guidelines Limit: Yes
Submission Guidelines Authorship: Yes
Student Author: Yes
Submission Number: 1191
Loading