Journal of Computer Applications ›› 2020, Vol. 40 ›› Issue (2): 454-458.DOI: 10.11772/j.issn.1001-9081.2019091665

• CCF NDBC 2019 • Previous Articles     Next Articles

Recommendation method with focus on long tail items

Jing QIN(), Qingbo ZHANG, Bin WANG   

  1. School of Computer Science and Engineering,Northeast University,Shenyang Liaoning 110169,China
  • Received:2019-08-12 Revised:2019-09-29 Accepted:2019-10-24 Online:2019-11-06 Published:2020-02-10
  • Contact: Jing QIN
  • About author:ZHANG Qingbo, born in 1994, Ph. D. candidate. Her research interests include recommendation system, artificial intelligence.
    WANG Bin, born in 1972, Ph. D., associate professor. His research interests include database, artificial intelligence.
  • Supported by:
    the National Natural Science Foundation of China(61572122)

关注长尾物品的推荐方法

秦婧(), 张青博, 王斌   

  1. 东北大学 计算机科学与工程学院,沈阳 110169
  • 通讯作者: 秦婧
  • 作者简介:张青博(1994—),女,河南三门峡人,博士研究生,主要研究方向:推荐系统、人工智能
    王斌(1972—),男,辽宁沈阳人,副教授,博士生导师,博士,主要研究方向:数据库、人工智能。
  • 基金资助:
    国家自然科学基金资助项目(61572122)

Abstract:

To solve the long tail problem caused by low coverage and diversity in recommendation system, a recommendation framework for long tail items and a recommendation algorithm named FLTI (Focusing on Long Tail Item) were proposed. The recommendation framework for long tail items was built based on Convolutional Neural Network (CNN) model with three layers, including data processing layer, recommendation algorithm layer and recommendation list generation layer. The FLTI algorithm was added to the recommendation algorithm layer of the framework, and calculated the frequent recommendation items and the infrequent recommendation items at first, then replaced the frequent recommendation items by the long-tail items to meet the specified proportion of long-tail items in the system. Experimental results on Movielens 1M and BookCrossing datasets show that compared to traditional User-Based Collaborative Filtering (UserCF) algorithm, Item-based Collaborative Filtering (ItemCF) algorithm, Singular Value Decomposition (SVD) recommendation algorithm and Colloborative Denosing Auto-Encoder (CDAE) algorithm, the coverage of FLTI algorithm is improved up to 51%, and the diversity of FLTI algorithm is improved up to 59% .

Key words: recommendation system, long tail recommendation, long tail item, frequent recommendation item, Convolutional Neural Network (CNN)

摘要:

针对推荐系统算法中覆盖率和多样性偏低所带来的长尾问题,提出了一种长尾物品的推荐框架以及关注长尾物品的推荐算法FLTI。长尾物品的推荐框架是基于卷积神经网络(CNN)模型构建的,分为数据处理层、推荐算法层和推荐列表生成层。将FLTI算法加入到了框架中的推荐算法层,该算法首先计算了频繁推荐项以及非频繁推荐项,然后采用使用长尾物品替换频繁推荐项的方法来满足系统中指定的长尾比例。实验结果表明,在Movielens 1M和BookCrossing数据集上,FLTI算法比传统的基于用户的协同过滤(UserCF)算法、基于物品的协同过滤(ItemCF)算法、奇异值分解(SVD)推荐算法以及协同去噪自动编码(CDAE)算法在覆盖率指标上最多提高了51%,多样性指标上最多提高了59%。

关键词: 推荐系统, 长尾推荐, 长尾物品, 频繁推荐项, 卷积神经网络

CLC Number: