多维框架下个性化推荐系统的流行度偏见测度方法与实证研究

Measuring Method and Empirical Study on Popularity Bias in Personalized Recommendation Systems under the Multidimensional Framework

[目的/意义]个性化推荐系统的流行度偏见表现为向用户推荐最热门而非最相关的物品,是影响公平性的重要因素之一,对推荐系统中的多个利益相关者产生严重影响。通过科学的定量方法准确识别流行度偏见,有助于评估现有推荐算法中存在的偏见问题,有助于优化系统,实现可信的人工智能,具有重要的理论和实践价值。[方法/过程]通过对比分析方法,对当前推荐系统领域内主流的流行度偏见测度指标进行对比,以此为基础构建一个推荐系统流行度偏见发现的多维指标框架,并进行实证分析。[结果/结论]从多样性、个性化、用户满意度、公平性、长期效益、整体表现六个维度阐明了个性化推荐系统中流行度偏见的测量方法,实证结果表明,该测度方法指标覆盖面广、测量准确性高、科学实践性强,具有一定的先进性。

[Purpose/significance]The popularity bias of the personalized recommendation system is to recommend the most popular items rather than the most relevant ones to users.It is one of the important factors affecting fairness and has serious consequences for multiple stakeholders in recommendation systems.Accurately identifying popularity bias through scientific quantitative methods is of great theoretical and practical value for assessing the bias problems existing in current recommendation algorithms,helping to optimize the system and realizing trustworthy artificial intelli⁃gence.[Method/process]Through the comparative analysis,the mainstream popularity bias measurement metrics in the current recommendation system field are compared.Based on this,a multidimensional measurement metrics frame⁃work for discovering popularity bias in recommendation systems was constructed,and an empirical analysis was carried out.[Result/conclusion]From the six dimensions including diversity,personalization,user satisfaction,fairness,longterm benefits,and overall performance,the measurement method of popularity bias in personalized recommendation systems is clarified.The empirical results show that the measurement method has a certain degree of advancement in terms of wide coverage of indicators,high measurement accuracy,and strong scientific practice.