The Debiasing Recommendation Algorithm Based on Pseudo-label Uncertainty Verification
Huiqiang Wu, Ruoning Song, Lan Ma, Zhenyu Liu, Jie Yang, Yuanyuan Qiao · 2024
Existing studies on recommender system mostly fit user data by building models, but these methods often lead to inconsistencies between offline and online indicators and may damage user experience because of various biases in the observed data.This paper focuses on the exposure bias and popularity bias problems in recommender systems, and proposes a debiasing recommendation algorithm based on pseudo-label uncertainty verification, aiming to reduce the impact of these biases and improve the quality of recommendations.By combining semi-supervised learning and Inverse Propensity Score (IPS) techniques, the algorithm can effectively identify potential positive samples in non-interaction samples and suppress exposure bias and popularity bias.In addition, this paper also proposes a pseudo-label uncertainty verification method based on self-adaptive threshold, which adaptively adjusts the confidence threshold according to the model learning state, making the algorithm more flexible and effective.The algorithm presented in this paper achieved significant improvements on the Amazon-book dataset: 15.4% in Recall@10 and 12.0% in NDCG@10, as well as 14.0% in Recall@20 and 12.4% in [email protected], on the MovieLens 1M dataset, the algorithm showed significant enhancements, with increases of 11.8% in Recall@10 and 17.6% in NDCG@10, as well as 6.9% in Recall@20 and 13.7% in [email protected] experimental results demonstrate that the algorithm effectively reduces bias and improves the recommendation quality of the recommender system.