Graph representation learning models have demonstrated great capability in many real-world applications. Nevertheless, prior research indicates that these models can learn biased representations leading to discriminatory outcomes. A few works have been proposed to mitigate the bias in graph representations. However, most existing works require exceptional time and computing resources for training and fine-tuning. To this end, we study the problem of efficient fair graph representation learning and propose a novel framework FairMILE. FairMILE is a multi-level paradigm that can efficiently learn graph representations while enforcing fairness and preserving utility. It can work in conjunction with any unsupervised embedding approach and accommodate various fairness constraints. Extensive experiments across different downstream tasks demonstrate that FairMILE significantly outperforms state-of-the-art baselines in terms of running time while achieving a superior trade-off between fairness and utility.
View on arXiv