主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2024, Vol. 32 ›› Issue (3): 20-27.doi: 10.16381/j.cnki.issn1003-207x.2021.1519

• • 上一篇    下一篇

基于“少数派报告”的协同过滤模型

黄晓宇1,2,鲜征征3(),杨雄文4,左文明1,2   

  1. 1.华南理工大学电子商务系, 广东 广州 510006
    2.人工智能与数字经济广东省实验室(广州), 广东 广州 510335
    3.广东金融学院互联网金融与信息工程学院, 广东 广州 510520
    4.华南理工大学法学院, 广东 广州 510006
  • 收稿日期:2021-09-03 修回日期:2022-02-22 出版日期:2024-03-25 发布日期:2024-03-25
  • 通讯作者: 鲜征征 E-mail:xianzhengzheng@126.com
  • 基金资助:
    广东省哲学社会科学规划项目(GD21CGL02);人工智能与数字经济广东省实验室(广州)青年学者项目(PZL2021KF0027);广东省自然科学基金项目(2020A1515010830);华南理工大学中央高校基本科研业务费项目(XYMS202107);广东省普通高校人文社会科学研究重点项目(2018WZDXM032);广州市科技计划项目(202002030473);广东省普通高校创新团队项目(2021KCXTD079)

Collaborative Filtering with Minority Report

Xiaoyu Huang1,2,Zhengzheng Xian3(),Xiongwen Yang4,Wenming Zuo1,2   

  1. 1.Department of Electronic Business, South China University of Technology, Guangzhou 510006, China
    2.Guangdong Artificial Intelligence and Digital Economy Laboratory (Guangzhou), Guangzhou 510335, China
    3.School of Internet Finance and Information Engineering, Guangdong University of Finance, Guangzhou 510520, China
    4.School of Law, South China University of Technology, Guangzhou 510006, China
  • Received:2021-09-03 Revised:2022-02-22 Online:2024-03-25 Published:2024-03-25
  • Contact: Zhengzheng Xian E-mail:xianzhengzheng@126.com

摘要:

协同过滤是在线推荐系统最重要的组成模块之一,为实现面向用户的个性化广告投放功能提供了关键的技术支持。然而,在协同过滤系统的具体实现中,研究者却不加区分地对所有目标用户都使用相同的训练流程。较之“个性化的预测结果”的预期,上述“一般化的训练过程”使得结果模型相对于待预测的目标过于普适,缺乏必要的针对性。本文提出一个“以个性化的训练过程得到个性化的预测结果”的协同过滤预测模型(MORE):对给定的目标用户,MORE将使用弹性网络模型(Elastic Net),从现有的用户全集中选出若干用户构成与之对应的 “专家”集合,并基于目标用户与专家已有的评分,生成对目标用户的缺失评分的预测。本文报告了MORE在不同的协同过滤预测模型上的应用结果,在真实评分数据集上的实验结果表明,较之使用全量数据训练得到的预测模型,基于MORE的模型有更好的表现。

关键词: 协同过滤, 稀疏选择, 个性化, 推荐系统

Abstract:

Collaborative filtering (CF for short) technology, as an important component of recommender system, plays a key role in personalized advertising. On the other hand, although various CF models are aimed at capturing personalized preferences of individuals, their training process are without regarding differences among the individual, where given a training algorithm, the model is always trained with the same training data, no matter what the target object is. This approach, however, is too general to the individuals, for given a specific target user, in order to personalize the recommendations, it might be more reasonable to first personalize the training process. In this work, MORE, a collaborative filtering model with MinOrity REport is presented. MORE trains the CF model in an user specific mode. Given a target user, MORE essentially works in three steps: First, it constitutes a personalized expert set for the target user with the Elastic Net model; Second, it constructs a sub matrix of the origin “user-item” rating matrix, where all values are only from the target user and the expert users. Third, it performs matrix imputation on the newly constructed matrix, and obtain estimations for the target user’s uncollected rating values. MORE is evaluated with two different CF models on two real movie rating data sets, all results show that the proposed model is promising.

Key words: collaborative filtering, sparse selection, personalization, recommender system

中图分类号: