基于特征权重与情感偏好的可解释推荐  

Explainable recommendation based on weighted feature and emotional preference

作　　者：戴兴 刘永坚[1] 解庆 刘平峰[2] DAI Xing;LIU Yong-jian;XIE Qing;LIU Ping-feng(School of Computer Science and Technology,Wuhan University of Technology,Wuhan 430070,China;School of Economics,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]武汉理工大学计算机科学与技术学院,湖北武汉430070 [2]武汉理工大学经济学院,湖北武汉430070

出　　处：《计算机工程与设计》2022年第8期2130-2136,共7页Computer Engineering and Design

基　　金：湖北省自然科学基金项目(2018CFB564);中央高校基本科研业务经费基金项目(2020III008GX)。

摘　　要：针对协同过滤算法在为用户商品相关性建模时未考虑用户/商品对特征属性的不同关注度及不可解释性问题,提出基于特征权重与情感偏好的可解释推荐算法。利用评论中抽取的特征及对应情感设计用户商品的表征,根据TF-IDF算法确定其重要性,将其加入相关性建模中;在评分预测时引入贝叶斯个性化排序减小评分误差;在生成推荐的同时,提供特征短语级别的解释。实验结果表明,对比现有模型,该模型均方根误差平均降低了3.62%,最大降低了4.93%。To solve the problem that the collaborative filtering algorithm does not consider the user/product’s different attention to feature attributes when modeling the user product relevance and inexplicability,an explainable recommendation algorithm based on weighted feature and emotional preference was proposed.The features and corresponding emotions extracted from reviews were used to design the representations of users and products.Their importance was determined according to the TF-IDF algorithm,and added to the correlation modeling.Bayesian personalized ranking was introduced to reduce rating errors when rating predictions.While generating recommendations,an explanation at the feature phrase-level was provided.Experimental results show that compared with the existing model,the root mean square error of the model is reduced by 3.62%on average,and the maximum is reduced by 4.93%.

关 键 词：可解释推荐 情感词典 特征权重 贝叶斯个性化排序 矩阵分解 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]