PBiLoss: Popularity-Aware Regularization to Improve Fairness in Graph-Based Recommender Systems
A model-agnostic loss function that boosts fairness without sacrificing accuracy.
Researchers Mohammad Naeimi and Mostafa Haghir Chehreghani have introduced PBiLoss, a novel regularization-based loss function designed to tackle popularity bias in graph-based recommender systems. Graph neural networks (GNNs) like LightGCN excel at modeling user-item interactions but often over-recommend popular items, leading to unfair exposure and reduced personalization. PBiLoss directly penalizes this tendency by augmenting traditional training objectives, encouraging the model to recommend less popular but more personalized content.
PBiLoss employs two innovative sampling strategies—Popular Positive (PopPos) and Popular Negative (PopNeg)—to distinguish popular items from niche ones. It offers flexibility through two methods: a fixed popularity threshold or a threshold-free approach, making it adaptable to different datasets. Extensive experiments on Epinions, iFashion, and MovieLens show PBiLoss reduces popularity bias metrics PRU and PRI by up to 10% compared to baselines, while preserving accuracy and other standard metrics. As a model-agnostic solution, it integrates seamlessly into existing GNN frameworks.
- PBiLoss uses PopPos and PopNeg sampling strategies to penalize popular item over-recommendation.
- Reduces popularity bias metrics PRU and PRI by up to 10% on Epinions, iFashion, and MovieLens.
- Model-agnostic and integrates with LightGCN and other GNN frameworks without accuracy loss.
Why It Matters
Fairer recommendations mean better user trust and discovery of niche content in e-commerce and media.