基于Co-training协同训练的在线虚假评论识别研究  

A study on deceptive review identification based on co-training

作　　者：张文 王强[1] 步超骐 李健 张思光 ZHANG Wen;WANG Qiang;BU Chaoqi;LI Jian;ZHANG Siguang(School of Economics and Management,Beijing University of Technology,Beijing 100124,China;School of Economics and Management,Beijing University of Chemical Technology,Beijing 100039,China;Institute of Science and Development,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]北京工业大学经济管理学院,北京100124 [2]北京化工大学经济管理学院,北京100039 [3]中国科学院科技战略咨询研究院,北京100190

出　　处：《系统工程理论与实践》2020年第10期2669-2683,共15页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金(71932002,71601023,61432001);西安市科技计划创新基金(文理专项,2016CXWL21)。

摘　　要：本文基于协同训练模型(co-training)提出了一种新的在线虚假评论识别方法CoDeRI以解决虚假评论识别中模型训练数据不足的问题.对同一评论信息,本文通过构建两个特征视图相互学习以识别虚假评论信息:视图一的特征来自于评论文本的词项(Term);视图二的特征来自于对评论进行深度语法树分析之后得到的概率上下文无关语法规则(PCFG,probabilistic context-free grammars).利用朴素贝叶斯(na?ve Bayes)作为基分类器,本文提出了两种特定于CoDeRI方法的分类后样本选择策略:CoDeRI-C策略和CoDeRI-U策略.CoDeRI-C策略在对未标注信息进行标注之后,选取分类置信度最高的评论信息以扩大训练样本集;CoDeRI-U策略则随机均匀的选取标注之后的评论信息以扩大训练样本集.实验表明,CoDeRI算法在虚假评论信息识别上与现有方法相比能够取得较好的分类结果.并且,CoDeRI-U策略的虚假评论识别准确率优于CoDeRI-C策略.本文的研究为电子商务中的平台、商家和消费者如何识别在线虚假评论提供了一定的管理启示.This paper proposes a co-training approach called CoDeRI to identify deceptive reviews with two independent views:One is trained with the lexical Terms and the other is trained with the PCFG(probabilistic context-free grammars)rules.By using naive Bayes as the base classifier,we develop two strategies as CoDeRI-C and CoDeRI-U embedded with the CoDeRI approach.The CoDeRI-C strategy uses unlabeled reviews classified with largest confidence to augment the training dataset to train the classifier iteratively.The CoDeRI-U strategy uses unlabeled reviews randomly with a uniform distribution on confidence to augment the training dataset to train the classifier iteratively.Experiments demonstrate that the proposed CoDeRI approach outperforms state-of-the-art techniques in deceptive review identification and the CoDeRI-U strategy outperforms the CoDeRI-C strategy.This study provides managerial implications for the platforms,merchants and consumers involved in e-commerce to identify online deceptive reviews.

关 键 词：协同训练 概率上下文无关文法 虚假评论 朴素贝叶斯 CoDeRI 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]