融合CatBoost的改进协同过滤推荐方法  

Improved Collaborative Filtering Recommendation Method Combined with CatBoost

作　　者：马冠勇 耿秀丽[1] 王海宇 华嘉颖 MA Guanyong;GENG Xiuli;WANG Haiyu;HUA Jiaying(Business School,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学管理学院,上海200093

出　　处：《软件导刊》2023年第7期21-26,共6页Software Guide

基　　金：国家自然科学基金项目(72271164);教育部人文社会科学研究规划基金项目(19YJA630021);高等学校博士学科点专项科研基金项目(20133120120002)。

摘　　要：传统协同过滤算法存在推荐精度不高、用户冷启动的问题。为解决以上问题,提出一种融合改进协同过滤与集成学习算法CatBoost的推荐方法,并讨论不同情况下推荐方法的多样性表现。该方法引入基于相对相似度指数(RSI)的重要最近邻(SNN)方法识别最近邻进行推荐预测;使用原始数据集训练CatBoost模型对改进协同过滤产生的召回集进行二次预测,并使用原始数据集对存在冷启动的用户进行预测,有效缓解冷启动问题;最后使用公开数据集Movie Lens进行比较实验,并讨论了在选择不同最近邻和TopK值时协同过滤算法的多样性表现。实验结果表明,改进协同过滤算法在准确率、召回率、F1 measure上均取得了更好的精度表现;融合CatBoost的改进协同过滤推荐方法虽相较融合随机森林和XGBoost的推荐方法有一定的多样性损失,但获得了更好的推荐精度。The traditional collaborative filtering method has the problems of low recommendation accuracy and cold startup of users,recom⁃mendation system which only considers accuracy can no longer meet the needs of current enterprises and users.To solve the above problems,a recommendation method combining improved collaborative filtering and the integrated learning algorithm CatBoost is presented,and the diver⁃sity of the recommendation methods in different situations is discussed.The significant nearest neighbors(SNN)based on relative similarity in⁃dex(RSI)are introduced,and the SNN method identifies the nearest neighbors for recommended prediction.Using the original dataset to train the CatBoost model,to make a second prediction of the recall generated by the improved collaborative filtering algorithm,and to make a pre⁃diction of the users who have cold start using the original dataset,can effectively alleviate the cold start problem.Finally,a comparative exper⁃iment is performed using an open dataset,Movie Lens,and the diversity of collaborative filtering algorithms when selecting different nearest neighbors and TopK values is discussed.The experimental results show that the improved collaborative filtering algorithm has achieved better precision performance than the traditional collaborative filtering algorithm in terms of Precision,Recall,and F1 measure.At the same time,the improved collaborative filtering recommendation method integrated with CatBoost has a certain diversity loss compared with the improved collaborative filtering recommendation method integrated with random forest and XGBoost,but it obtains better recommendation accuracy.

关 键 词：推荐方法 协同过滤 CatBoost 聚合多样性 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]