Multi-teacher knowledge distillation for debiasing recommendation with uniform data

Recent studies have highlighted the bias problem in recommender systems which affects the learning of users' true preferences. One significant reason for bias is that the training data is missing not at random (MNAR). While existing approaches have demonstrated the usefulness of uniform data that is missing at random (MAR) for debiasing, the current models lack a comprehensive exploration of unbiased features within uniform data. Considering the valuableness and limited size of uniform data, this paper proposes a multi-teacher knowledge distillation framework (UKDRec) to extract and transfer more unbiased information from uniform data. The proposed framework consists of two components: a label-based teacher model that leverages supervision signals and a feature-based teacher model that facilitates the transfer of comprehensive unbiased features. To effectively extract unbiased features, we introduce a contrastive learning strategy that combines the uniform data with control data. The framework is trained using a multi-task learning approach, which enhances the transfer of unbiased knowledge. Extensive experiments conducted on real-world datasets demonstrate the superior debiasing performance of our approach compared to competitive baselines.