Bias-based Denoising Causal Recommendation Algorithm

Abstract

Traditional recommendation algorithms, such as collaborative filtering, make recommendations by learning the relevant relationships between users and items. However, considering only the relationships without considering the underlying causal mechanisms would be unfair, uninterpretable, and would lead to bias. In this paper, we propose bias-based denoising causal recommendation algorithm (BDCR) . First, the method dynamically transforms the explicit user-item feedback into implicit feedback with an embedded representation. Then, a truncation function based on causal inference is constructed to remove false positive noise. In addition, traditional recommendations and denoised causal recommendations are aggregated to obtain predictive scores. Finally, experimental results on two real datasets show that the BDCR algorithm outperforms the classical algorithm in terms of recall and NDCG metrics.