Fair Representation Learning for Heterogeneous Information Networks

Introduction

• Demographic bias in HIN representations has received little attention, despite the risk that it may cause unfairness in downstream applications.The paper notes that systems trained on human-generated content can inherit or amplify biases in that data.Source: Introduction · p. 1
• The paper proposes sampling-based, projection-based, and GNN-based algorithms to mitigate demographic bias in HIN representations.It systematically studies how the methods balance prediction accuracy and fairness.Source: Introduction · pp. 1–2
• The evaluation applies the proposed methods to automated career counseling, where biased advice may restrict opportunities for disadvantaged groups.The paper highlights girls and minorities as examples of affected populations.Source: Introduction · pp. 1–2
• The authors present this work as the first systematic investigation of measuring and mitigating demographic bias in HIN representations.They distinguish earlier fair network-embedding work as focused on homogeneous networks.Source: Introduction · p. 2
• Experiments on two real-world datasets demonstrate mitigation of gender bias in career recommendation and examine accuracy–fairness trade-offs.The results are intended to guide practitioners in choosing methods under different conditions.Source: Introduction · p. 2

Related Work

• The paper frames its subject as HIN representation learning and fairness methods, building on representation-learning research and broader efforts to address societal harm from machine-learning bias.The surveyed HIN methods include both general-purpose embeddings and end-to-end supervised GNN approaches.Source: Related Work · p. 2, HIN Representation Learning · p. 2, Task-specific HIN Representation… · p. 2

HIN Representation Learning

• Task-agnostic HIN methods commonly use meta-path-guided sampling followed by word-embedding algorithms, while task-specific methods optimize a downstream task.A meta-path is a sequence of relations across node types, such as author–paper–author.Source: HIN Representation Learning · p. 2, Task-agnostic HIN Representation… · p. 2

Fair Machine Learning: General Approaches

• Fair machine-learning approaches include preprocessing, fairness-guided optimization, and postprocessing of model predictions.Preprocessing may learn representations independent of protected attributes; optimization adds fairness penalties or constraints.Source: Fair Machine Learning: General… · p. 2

Fair Word Embeddings

• Word embeddings can inherit or amplify societal stereotypes in text, and common debiasing projects embeddings orthogonally to a bias direction.An alternative computes the demographic-bias direction using gendered person names.Source: Fair Word Embeddings · p. 3

Fair Homogeneous Network Embedding

• Prior fair network-embedding methods include fairness-aware random walks for homogeneous networks and adversarial graph embeddings for link prediction.The paper says the former is limited to homogeneous networks, while the latter does not use HIN properties to learn embeddings and enforce fairness.Source: Fair Homogeneous Network Embedding · p. 3

Problem Statement

• The setting is an HIN with multiple node types, edges, node types, and relations, plus a binary protected attribute associated with at least one node type.Examples include users, careers, and items, with relations such as liking a page or choosing a career.Source: Problem Statement · p. 3
• The goal is to learn node embeddings that support accurate held-out-edge link prediction while behaving fairly with respect to protected demographics.Career counseling illustrates the task: predicting career choices from users’ interests, such as liked pages.Source: Problem Statement · p. 3
• Demographic parity requires predicted-label distributions to be the same across protected groups for each class label.For career recommendation, this means each career has the same recommendation probability for male and female users.Source: Problem Statement · p. 3
• Equal opportunity requires equal prediction rates across protected groups among examples whose ground-truth label is the class being predicted.For career recommendation, compare recommendation rates for a career among users who actually chose that career.Source: Problem Statement · p. 4

Methodology

• The paper presents fairness-aware sampling, embedding projection, and GNNs with fairness-aware losses as three classes of fair HIN representation methods.The methods are illustrated through fair career counseling, formulated as fair link prediction in HINs.Source: Methodology · p. 4

Fairness-aware Sampling

• Fair meta-path sampling increases the chance of sampling users from the disadvantaged group and decreases it for the advantaged group.Sampling probabilities are inversely proportional to group size; the career-to-user transition uses a hyperparameter to tune how strongly the walker favors a group. M2V learns embeddings from meta-path walks, and an MLP predicts careers from user and career embeddings.Source: Fair Meta-path Generation · pp. 4–5, Fairness-aware Sampling · p. 4

Embedding Projection

• Embedding projection reduces bias by projecting each user embedding orthogonally to a bias direction derived from group embeddings.For a binary protected attribute, the group directions are computed by averaging users’ embeddings within each group.Source: Embedding Projection · p. 5

Fairness-aware Graph Neural Networks (GNNs)-based Representation Learning

• GNNs frame career recommendation as user-node classification and add a fairness-aware loss alongside the accuracy objective.The paper considers demographic parity and equal opportunity; probabilistic predictions replace hard counts because counts do not support gradient backpropagation.Source: Fairness-aware Graph Neural Networks… · p. 5, Fairness-aware Loss · pp. 5–6

Experiments

• The experiments test the methods on Facebook and MovieLens and compare their accuracy–fairness trade-offs using Pareto frontiers and fixed fairness thresholds.The frontier plots fairness against MRR; the threshold analysis selects the highest-MRR model satisfying each constraint.Source: Experiments · p. 6, Experimental Results · p. 8

Datasets

• Facebook has 7,069 users, 99,756 page nodes, and 48 career nodes; MovieLens has 4,920 users, 3,677 movie nodes, and 14 career nodes after career filtering.Users connect to pages or movies through likes or ratings and to careers through declared career links.Source: Datasets · p. 6

Evaluation Metrics

• The evaluation measures career-prediction accuracy with mean reciprocal rank and fairness with demographic parity and equal opportunity.The result tables report mean MRR over five runs under low-, medium-, and high-fairness conditions.Source: Evaluation Metrics · p. 7, Table 2 · p. 9, Table 3 · p. 9

Compared Methods

• Eight methods include non-debiased M2V and GNN baselines, balance-data and adversarial baselines, and proposed methods combining sampling, projection, or fairness-aware GNN losses.Sampling-based methods use learned user and career embeddings as MLP inputs for career ranking.Source: Compared Methods · p. 7

Model Selection

• The authors tune method-specific hyperparameters, including walk settings and sampling ratio, fairness-loss weights, and the adversarial accuracy–fairness trade-off coefficient.Grid search is used for sampling, GNN, and adversarial methods.Source: Model Selection · p. 8

Experimental Results

• On Facebook, fair-sampling methods perform well at medium fairness, while GNN-demographic-parity and GNN-equal-opportunity perform well under high fairness; MovieLens favors projection-based methods at low and medium fairness and fairness-aware GNNs at high fairness.For Facebook, GNN-equal-opportunity performs best under all equal-opportunity conditions; for MovieLens, the high-fairness GNN methods are comparable with the adversarial method.Source: Experimental Results · pp. 8–9, Table 2 · p. 9, Table 3 · p. 9

• The study investigates sampling-, projection-, and GNN-based approaches to debiasing HIN representation learning.Source: Conclusion · p. 9
• The methods are evaluated for mitigating gender bias in career recommendation across two datasets.Source: Conclusion · p. 9
• Different algorithms are effective under different conditions, providing guidance to practitioners; community recommendation is left for future work.Source: Conclusion · p. 9