FairExpand: Individual Fairness on Graphs with Partial Similarity Information

TL;DR

FairExpand tackles graph-based individual fairness when full similarity information is unavailable. It combines a backbone model that enforces partial fairness using observed similarities with an auxiliary PU-learning-based link predictor (PULL) to expand similarity information via epsilon-greedy exploration, iteratively propagating fairness constraints across the graph. Empirical results across six real-world graphs show FairExpand consistently improves fairness (lower Bias) while maintaining competitive task performance (F1), outperforming several full-information baselines under partial data. The method is adaptable to different backbones and tasks, making graph-based individual fairness practical in real-world settings with limited annotations.

Abstract

Individual fairness, which requires that similar individuals should be treated similarly by algorithmic systems, has become a central principle in fair machine learning. Individual fairness has garnered traction in graph representation learning due to its practical importance in high-stakes Web areas such as user modeling, recommender systems, and search. However, existing methods assume the existence of predefined similarity information over all node pairs, an often unrealistic requirement that prevents their operationalization in practice. In this paper, we assume the similarity information is only available for a limited subset of node pairs and introduce FairExpand, a flexible framework that promotes individual fairness in this more realistic partial information scenario. FairExpand follows a two-step pipeline that alternates between refining node representations using a backbone model (e.g., a graph neural network) and gradually propagating similarity information, which allows fairness enforcement to effectively expand to the entire graph. Extensive experiments show that FairExpand consistently enhances individual fairness while preserving performance, making it a practical solution for enabling graph-based individual fairness in real-world applications with partial similarity information.

Figures (5)

• Figure 1: Illustration of FairExpand for individually fair graph representation learning with partial similarity information. This framework operates as an iterative pipeline with two components: (1) a partial fairness enforcement phase, where individual fairness is enforced on a backbone GNN model using the observed similarity data $S^{(0)}$, and (2) a similarity propagation phase, which uses an auxiliary link predictor and exploration techniques to gradually expand similarity information to larger regions of the graph. By alternating between these phases, FairExpand effectively scales fairness enforcement across the entire graph.
• Figure 2: Sensitivity analysis of F1 and bias scores with respect to the number of similarities in $S^{(0)}$, denoted $|S^{(0)}|$, and the number of edges $m_{\text{add}}$ added in Phase 2 of FairExpand.
• Figure 3: Exploration trajectories after a Phase 2 iteration in FairExpand ablations. Larger nodes indicate original members of $S^{(0)}$. We observe that $\varepsilon$-greedy PULL (red edges) exhibits the broadest exploration patterns.
• Figure 4: Distributions of pairwise cosine similarities $S_{ij}$.
• Figure 5: Bias trajectories over FairExpand iterations.