Debiased Recommendation Model Based on User Historical Behavior
Liliang Tao, Yurong Song, Hong-Bo Qu, Ruqi Li · 2023
With the rapid development of the Internet, there is a widespread problem of highly imbalanced long tail distribution of user project interactions in most real-world application recommendation scenarios. This paper presents a novel approach for mitigating bias in the recommendation process of recommendation system models through the introduction of a debiased recommendation model that leverages user history behavior. The model models user click behavior on recommendation results and utilizes graph convolution operations to embed user and item representations, while incorporating recommendation ranking position information and item popularity information into the prediction inputs. Additionally, false positive signals are employed as post-processing to refine the model's prediction scores, resulting in improved item recommendation accuracy and enhanced personalized recommendation capabilities. The experimental results demonstrate that the proposed model exhibits a notable improvement in both precision and accuracy when compared to other baseline models mentioned in the paper.