基于改进的反事实推理框架消除流行度偏差

Elimination of Popularity Bias Based on Improved Counterfactual Inference Model

本文将时间因子融入反事实推理框架中,得到改进的无关模型的反事实推理框架。推荐系统的最终目的是提供给用户个性化的建议,而不是推荐流行的项目。借助因果关系中的反事实推理法来分析解决流行度偏见问题是一个新颖而基本的视角。流行度偏差是指热门的项目展示相对较多,系统推荐的时候,会倾向于推荐热门商品。当前的推荐模型只是考虑用户项目交互对最终评分的影响,忽略了时间因子对于交互过程的影响,所以对于流行度偏差消除的结果并不理想。由于用户和项目的匹配度不是一成不变的,它会随着时间非线性变化,为了提升模型消除流行度偏差的能力,本论文将时间因子纳入到了遵循因果推理的反事实推理框架中,从用户和项目交互的细节出发,细化了用户项目匹配,完成了对反事实推理框架的改进。实验使用矩阵分解算法和MovieLens数据集进行实验,融入时间因子的反事实推理框架对比传统的框架,命中率、召回率、归一化折损累计增益都有了一定的提高,提高了推荐算法的去流行度偏差能力。

This paper integrates the time factor into the counterfactual reasoning framework and obtains an improved model-independent counterfactual reasoning framework. The ultimate purpose of the recommendation system is to provide users with personalized recommendations, rather than rec-ommending popular items. It is a novel and fundamental perspective to analyze and solve the problem of popularity bias with the help of counterfactual reasoning in causal relationships. Popularity bias means that relatively more popular items are displayed, and when the system recom-mends them, it will tend to recommend popular items. The current recommendation model only considers the impact of user item interaction on the final score and ignores the impact of the time factor on the interaction process, so the results for eliminating the popularity bias are not ideal. Since the matching degree between users and items is not static, it will change non-linearly with time. In order to improve the model’s ability to eliminate popularity bias, this paper incorporates the time factor into the counterfactual reasoning framework that follows causal reasoning, from users and Starting from the details of item interaction, user-item matching is refined and the improvement of the counterfactual reasoning framework is completed. The experiment uses the ma-trix decomposition algorithm and the MovieLens data set. Compared with the traditional frame-work, the counterfactual reasoning framework that incorporates the time factor has a certain improvement in hit rate, recall rate, and normalized loss cumulative gain, which improves the popu-larity bias removal ability of the recommendation algorithm.