[Paper Review] Fairness-Aware Node Representation Learning

Öykü Deniz Köse, Yanning Shen|arXiv (Cornell University)|Jun 9, 2021

Advanced Graph Neural Networks51 references10 citations

TL;DR

This paper proposes fairness-aware graph augmentations for contrastive node representation learning, using adaptive feature masking and edge deletion guided by fairness notions to reduce bias. It demonstrates improved fairness in statistical parity and equal opportunity—without sacrificing classification accuracy—on real social networks, making it the first adaptive, fairness-aware graph augmentation framework for contrastive learning.

ABSTRACT

Node representation learning has demonstrated its effectiveness for various applications on graphs. Particularly, recent developments in contrastive learning have led to promising results in unsupervised node representation learning for a number of tasks. Despite the success of graph contrastive learning and consequent growing interest, fairness is largely under-explored in the field. To this end, this study addresses fairness issues in graph contrastive learning with fairness-aware graph augmentation designs, through adaptive feature masking and edge deletion. In the study, different fairness notions on graphs are introduced, which serve as guidelines for the proposed graph augmentations. Furthermore, theoretical analysis is provided to quantitatively prove that the proposed feature masking approach can reduce intrinsic bias. Experimental results on real social networks are presented to demonstrate that the proposed augmentations can enhance fairness in terms of statistical parity and equal opportunity, while providing comparable classification accuracy to state-of-the-art contrastive methods for node classification.

Motivation & Objective

• To address the lack of fairness considerations in graph contrastive learning, where models may amplify biases in graph structure and features.
• To develop adaptive graph augmentation schemes that reduce intrinsic bias by targeting sensitive attributes in nodal features and topology.
• To introduce novel fairness concepts on graphs that guide the design of bias-reducing augmentations.
• To theoretically and empirically validate that adaptive feature masking reduces correlation between features and sensitive attributes.
• To achieve improved fairness metrics (statistical parity, equal opportunity) comparable to state-of-the-art contrastive methods in node classification.

Proposed method

• Proposes fairness-aware graph augmentations that adaptively corrupt features and edges based on node centrality and structural roles.
• Introduces two adaptive edge deletion strategies: degree-aware and triangle-based (based on local clustering), to reduce structural bias.
• Applies adaptive feature masking that selectively masks features based on their correlation with sensitive attributes, reducing expected correlation.
• Theoretical analysis proves that adaptive feature masking reduces the expected total sample correlation coefficient (ρ) between features and sensitive attributes compared to non-adaptive masking.
• Uses contrastive learning with two augmented views per node, where augmentations are fairness-guided and computationally lightweight.
• Employs a self-supervised contrastive objective to train node representations while minimizing bias propagation through adaptive augmentation design.

Experimental results

Research questions

• RQ1 Can adaptive graph augmentations reduce bias in node representations without degrading performance in node classification tasks?
• RQ2 How does adaptive feature masking compare to non-adaptive masking in reducing correlation between nodal features and sensitive attributes?
• RQ3 Can fairness-aware edge deletion strategies improve fairness metrics such as statistical parity and equal opportunity?
• RQ4 Do fairness-aware augmentations outperform existing contrastive learning and fairness-aware baselines in fairness while maintaining comparable accuracy?
• RQ5 How do different fairness notions on graphs (e.g., statistical parity, equal opportunity) inform the design of effective augmentations?

Key findings

• The proposed FM + p^(deg) and FM + p^(T) augmentations achieve the lowest fairness gaps: ΔSP = 2.50% and ΔEO = 2.18% on Pokec-z, outperforming all baselines.
• On Pokec-n, FM + p^(deg) achieves ΔSP = 4.39% and ΔEO = 6.05%, showing consistent fairness improvement across datasets.
• The adaptive feature masking scheme reduces the expected correlation between nodal features and sensitive attributes, as proven theoretically.
• Despite improving fairness, the proposed methods maintain classification accuracy comparable to state-of-the-art contrastive methods (e.g., GCA, GRACE), with FM + p^(deg) achieving 66.42% accuracy on Pokec-n.
• NIFTY, a fairness-aware method focused on counterfactual fairness, underperforms in statistical parity and equal opportunity, indicating that fairness goals must be aligned with specific fairness notions.
• All contrastive learning baselines outperform random-walk-based methods (e.g., FairWalk), confirming that graph topology propagates bias, which the proposed augmentations effectively mitigate.