基于直觉模糊集的伯努利矩阵分解推荐算法  

Intuitionistic fuzzy sets based Bernoulli matrix factorization recommendation algorithm

作　　者：邓江洲 郭均鹏[1] DENG Jiang-zhou;GUO Jun-pengy(College of Management and Economics,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学管理与经济学部,天津300072

出　　处：《控制与决策》2023年第10期2897-2904,共8页Control and Decision

基　　金：国家自然科学基金项目(72171165);教育部人文社科研究项目(21YJA630021);天津市哲学社会科学研究规划项目(TJGL17-011)。

摘　　要：现有的基于矩阵分解的协同过滤推荐算法主要从定量的角度,利用用户的评分信息评估模型表现,而并未从定性的角度描述用户的不确定偏好信息.鉴于此,从用户偏好模糊概率的角度提出一种基于直觉模糊集的伯努利矩阵分解推荐算法为目标用户进行Top-n推荐.首先,根据用户偏好特征和直觉模糊集定义,将用户评分矩阵划分为隶属度矩阵、非隶属度矩阵和犹豫度矩阵;然后,借助伯努利矩阵分解模型对矩阵并行拟合,得到最优的潜在特征向量对,并将其内积按比例划分,从而获得目标用户对未评分项目偏好程度的直觉模糊数;最后,根据直觉模糊数排序规则确定最终推荐列表.在公开数据集上的实验结果显示,所提出方法在项目排序指标上均优于其对比方法,能够有效提高推荐质量.Existing matrix factorization based collaborative filtering recommendation algorithms mainly utilize users’ratings to evaluate model performance from a quantitative perspective,and never describe users’uncertain preference information from qualitative perspective.Therefore,this paper proposes a Bernoulli matrix factorization recommendation model based on intuitionistic fuzzy sets(IFSs)to make Top-n recommendations for active users from the perspective of fuzzy probability of user preferences.Firstly,the user-item rating matrix is divided into the membership matrix,nonmembership matrix and hesitancy matrix according to user preference features and the definition of IFS.Subsequently,the Bernoulli matrix factorization(BeMF)is adopted to fit these matrices in parallel to obtain the optimal latent feature vectors,and their inner products are divided proportionally to get the intuitionistic fuzzy number(IFV)of the active users’preference degree for unrated items.Finally,the recommendation lists are determined according to the ranking rule of the IFV.Experimental results on several benchmark datasets show that the proposed model outperforms other methods in terms of item ranking metrics and effectively improves the recommendation quality.

关 键 词：直觉模糊集 伯努利分布 矩阵分解 推荐算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]