PBiLoss: A Regularization Method to Reduce Popularity Bias in Graph-Based Recommenders
Researchers propose PBiLoss, a regularization-based loss function to counteract popularity bias in graph neural network (GNN)-based recommender systems. Popularity bias leads to over-recommendation of popular items, reducing personalization, unfair exposure, and lowering diversity. Existing solutions include pre-processing (distorts data), in-processing (complicates optimization), and post-processing (limited in correcting embedded bias). PBiLoss augments traditional training by penalizing model inclination toward popular items. The paper is available on arXiv (2507.19067).
Key facts
- PBiLoss is a regularization-based loss function.
- It targets popularity bias in GNN-based recommenders.
- Popularity bias causes over-recommendation of popular items.
- It reduces personalization and fairness.
- Existing methods: pre-processing, in-processing, post-processing.
- Pre-processing can distort data distributions.
- In-processing complicates optimization.
- Post-processing is limited in correcting embedded bias.
- PBiLoss penalizes model inclination toward popular items.
- Paper available on arXiv (2507.19067).
Entities
Institutions
- arXiv