Recommendation Unlearning Algorithm Combining Fuzzy Clustering and Adaptive Denoising

Abstract

Privacy protection plays a crucial role in recommender systems as it helps to protect users’ sensitive information from disclosure risks. Recent recommendation unlearning has attracted increasing attention as an effective method of privacy protection. Existing methods often partition data into sub-partitions before training to enhance model training efficiency. However, simply partitioning interactions into sub-partitions can disrupt the integrity of user-item relationships and reduce the availability of data. In addition, the presence of false-positive noise in sub-partitions with implicit feedback can interfere with model training, preventing it from accurately capturing users’ true preferences. To address these challenges, a recommendation unlearning algorithm combining fuzzy clustering and adaptive denoising (FDRU) is proposed. Firstly, fuzzy clustering determines membership by calculating cosine distances between samples and various cluster centers, subsequently dividing the training dataset into several sub-partitions. Then, FDRU designs an adaptive denoising algorithm that dynamically eliminates false positive noise in sub-partitions based on thresholds. Finally, it utilizes dynamic weighted aggregation of sub-models for prediction and top-N recommendations. In order to assess the performance of the proposed algorithm, extensive experiments are carried out on three public datasets. Experimental results indicate that FDRU outperforms other benchmark algorithms on Recall and NDCG.

Keywords

• privacy protection; recommendation; unlearning; fuzzy clustering; adaptive denoising