基于动态多模网络的虚假评论检测方法研究

发布时间：2018-12-12 21:43

【摘要】：web2.0技术的迅速崛起,使越来越多的用户喜欢在电商平台和点评网站上发表评论,分享他们对于产品和服务的观点和感受,这些用户发布的评论信息无论是对消费者还是商家都是至关重要的,因为这些评论包含着大量用户对产品或者服务质量的描述。但是受利益的驱使,一些不法商家通过雇佣虚假评论者发布不真实的评论来提高自己的信誉或者诋毁竞争对手的信誉,以达到误导消费者购物决策的目的。这种行为不仅误导消费者的购物决策,而且还严重影响了电子商务的健康发展,所以尽早发现虚假评论并在最大程度上减少它们的影响是刻不容缓的。近年来,虚假评论检测已经成为一个热门的研究领域。研究者常常通过分析文本极性和评分模式来发现虚假攻击,这些通用的检测方法能够轻松地检测出常规的虚假攻击,但是却很难有效识别出那些把自己伪装成真实用户的虚假评论者。传统的单一维度检测算法未能考虑多个评论特征之间的潜在影响,致使准确率不高,为此本文提出了一种基于动态多模网络的虚假评论检测算法,并进行了较为深入的研究工作。本文主要工作及创新点如下:(1)提出了一种融合动态多模网络的虚假评论探测方法。该方法首先构建了包含评论、评论者、商品和商家的四维网络;然后提出了评论忠实度、评论者信誉度、商品优质度和商家可信度概念并对其量化;紧接着使用谱聚类算法探讨了四类节点之间的联系,最后设计了一个迭代计算模型,通过迭代计算揭示了四维网络之间的动态交互影响。使用该方法可以同时准确地检测出虚假评论、虚假评论者和不良商家。(2)提出了一种基于情感强度的虚假评论检测算法,该方法主要通过自然语言处理技术分析评论文本情感极性。在本文中,我们的方法主要有以下几点创新:首先,我们使用领域词典挖掘出评论类别,并考虑了关联词对文本极性的影响;其次,本文简化了实验数据的采集与处理工作,通过分析数据发现了5个重要的虚假评论检测特征;最后,使用逻辑回归模型将5个量化后的特征融合在一起,并训练出一个有效的虚假评论分类模型。该方法是计算多模网络中评论忠实度的重要前提。(3)提出了一种改进的基于用户信誉的虚假评论检测算法。首先,使用矩阵补全理论把低秩稀疏的用户-项目评分矩阵填充,其次,构建用户信誉评估模型;最后,本文选择了更加合理的预估标准,并且细化了群组规模相同而评分不同的用户信誉,使用top-k算法判定信誉值最低的k个用户为虚假评论者。该方法对于计算多模网络的用户信誉是至关重要的。
[Abstract]:The rapid rise of web2.0 technology has made more and more users like to comment on e-commerce platforms and comment sites to share their views and feelings about products and services. Comments posted by these users are critical to both consumers and businesses because they contain a large number of user descriptions of product or service quality. However, driven by interests, some illegal businesses use false reviewers to release false comments to improve their credibility or discredit competitors, in order to mislead consumers to make shopping decisions. This behavior not only misleads consumers' shopping decisions, but also seriously affects the healthy development of electronic commerce, so it is urgent to find false comments as soon as possible and minimize their impact. In recent years, false comment detection has become a hot research field. Researchers often detect false attacks by analyzing text polarity and scoring patterns, which can easily detect conventional false attacks. But it's hard to identify false commentators who pretend to be real users. The traditional single dimensional detection algorithm fails to take into account the potential influence of multiple comment features, which leads to low accuracy. Therefore, this paper proposes a false comment detection algorithm based on dynamic multi-mode network. And has carried on the more thorough research work. The main work and innovations of this paper are as follows: (1) A new method of false comment detection based on dynamic multimode network is proposed. In this method, a four-dimensional network including comments, reviewers, commodities and merchants is constructed, and then the concepts of comment fidelity, commenters reputation, commodity quality and merchant credibility are proposed and quantified. Then the relationship between the four kinds of nodes is discussed by using spectral clustering algorithm. At last, an iterative computing model is designed to reveal the dynamic interaction between the four dimensional networks. Using this method, false comments, false reviewers and bad merchants can be detected accurately at the same time. (2) A false comment detection algorithm based on emotional intensity is proposed. This method mainly uses natural language processing technology to analyze the emotional polarity of comment text. In this paper, our method mainly has the following innovations: first, we use the domain dictionary to mine out the comment categories, and consider the influence of the relevance words on the polarity of the text; Secondly, this paper simplifies the collection and processing of experimental data, and finds five important features of false comment detection by analyzing the data. Finally, the five quantized features are fused by using the logical regression model, and an effective false comment classification model is trained. This method is an important prerequisite for computing the fidelity of comments in multimode networks. (3) an improved algorithm for detecting false comments based on user reputation is proposed. Firstly, the matrix complement theory is used to fill the low rank sparse user-item scoring matrix. Secondly, the user reputation evaluation model is constructed. Finally, we select more reasonable prediction criteria, and refine the reputation of users with the same group size and different ratings. We use top-k algorithm to determine k users with the lowest reputation as false reviewers. This method is very important for computing the user reputation of multimode network.
【学位授予单位】：山东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1

【参考文献】