ACM Transactions on Information Systems (TOIS) / Recommender Systems

Mitigating Popularity Bias in Recommendation with Global Listwise Learning and Progressive Bi-Weighting

An accepted ACM Transactions on Information Systems paper studying popularity-bias mitigation through global listwise learning and progressive bi-weighting.

Publication record
Infographic for Mitigating Popularity Bias in Recommendation with Global Listwise Learning and Progressive Bi-Weighting

Research area

User reactivation, generative recommendation, Semantic-ID interface diagnostics, dynamic retrieval, graph embedding, counterfactual watch-time prediction, and efficient CTR modeling form the main technical line.

Publication record

Authors
Tianyu Zhu, Jiandong Ding, Yansong Shi, Guoqing Chen, Jian-Yun Nie
Venue
ACM Transactions on Information Systems (TOIS)
Area
Recommender Systems

Problem

Listwise recommendation objectives can be sensitive to misspecified propensities, while aggressive rebalancing can adversely affect representation learning.

Approach

The paper studies global listwise learning under inverse propensity scoring and introduces bi-weighting that smooths propensity estimates with a collection model. Progressive Bi-Weighting is designed to reduce the adverse effects of aggressive rebalancing on learned representations.

Analysis and evaluation

The work analyzes the collection model used for weighting and evaluates the proposed methods against established and recent debiasing baselines.