引用本文:陈晋音,吴洋洋,林翔.基于图过滤的快速密度聚类双层网络推荐算法[J].控制理论与应用,2019,36(4):542~552.[点击复制]
CHEN Jin-yin,WU Yang-yang,LIN Xiang.Double layered recommendation algorithm based on fast density clustering with graph-based filtering & Applications[J].Control Theory and Technology,2019,36(4):542~552.[点击复制]
基于图过滤的快速密度聚类双层网络推荐算法
Double layered recommendation algorithm based on fast density clustering with graph-based filtering & Applications
摘要点击 2216  全文点击 1111  投稿时间：2017-11-10  修订日期：2018-05-08
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2018.70816
  2019,36(4):542-552
中文关键词  对抗生成式网络,自动点评,基于图的过滤器,聚类推荐算法
英文关键词  Generative  Adversarial Nets,Automatic  Reviewer, Graph-Based  Filter, Clustering  based Recommender  algorithm
基金项目  国家自然科学基金(61502423)
作者单位E-mail
陈晋音* 浙江工业大学 chenjinyin@zjut.edu.cn 
吴洋洋 浙江工业大学  
林翔 浙江工业大学  
中文摘要
      信息过载问题使得推荐系统迅速发展并广泛应用,同时也出现不法商家将虚假消费记录定量地输入到系统数据库从而改变推荐系统的推荐结果以获利。因此,本文围绕三个问题展开,即：为了提高推荐系统对虚假评论的鉴别能力,首先需要准确标注虚假评论的类标,如何能获取大量准确标定的虚假评论信息；如何有效过滤虚假评论从而提高推荐的可靠性；如何实现一种高效可靠的推荐系统。针对虚假评论信息难以准确标定,本文提出了一种基于文本生成式对抗网络的自动点评技术,依据历史评论文本自动生成模拟评论文本,并依据情感分析确定生成文本的对应评分；为提高推荐系统对虚假评论的过滤的能力,提出一种能快速确定节点执行度阈值的基于图的过滤器；最后本文设计了基于图结构过滤的快速密度聚类双层网络,提高推荐的准确率。将所提出的推荐算法应用到Yelp数据集上展开实验,验证过提出方法的有效性。
英文摘要
      The information overloading problem leads to wider application of recommender system. At the meantime, fake reviewers are quantitative input into the history review records by illegal business to affect the recommender to change for their benefits. Three research questions are addressed in our paper. In order to improve fake review filtering ability for recommenders, abundant of accurately labeled fake reviewers are necessary. How to collect large amount of accurately labeled fake reviewers? How to filter fake reviewers accurately and efficiently? How to design an efficient recommender? Since it’s difficult to collect labeled fake reviewers, an automatic reviewer generator based on text generative adversarial nets (TextGAN) is proposed. Reviewers labeled as fake can be generated based on historical reviewers and can be rated according to emotional analysis. Aiming at improving filtering ability of recommender, we design a graph-based filter that can quickly determine node execution thresholds. And a recommender based on fast clustering is put forward, which is a density based clustering algorithm with cluster center self-determined, to implement accurate recommendation. At last, the proposed algorithm is applied to the Yelp data set to verify its effectiveness.