用户信誉度与用户生成内容质量评估模型研究

发布时间：2018-05-24 05:51

本文选题：户信誉度 + 用户生成内容质量　；参考：《清华大学》2014年硕士论文

【摘要】：Web2.0的到来促进了户成内容的繁荣发展，在促进知识传播与解决信息不对等的同时，同时也产了户恶意破坏他编辑的内容、伪造虚假评论、不尊重版权上传他成果等问题。这些问题的产使得对户信誉度与户成内容质量的评估成为必须。本先阐述了数据的获取、预处理与特征的抽取。我们以英维基百科作为知识分享类站的代表进研究，下载了英维基百科的数据，针对维基百科的章与作者的特点，我们抽取了量特征。我们随机抽取了3个类作为三个数据集，以数据集下的章与章的作者作为研究对象，并对章质量与作者的信誉度进标注，得到标准数据集。我们以亚马逊评论作为商务评论站的代表进研究，下载了亚马逊的评论数据，并随机的选取电产品下的评论与评论者作为研究对象，先对数据集进预处理解决数据稀疏性问题，针对亚马逊评论和评论者的特点，，抽取了量特征，并通过标注对评论与评论者进标注，得到标准数据集。在维基百科中，我们提出了双翼因图模型，该模型利我们所抽取的特征集，将作者的信誉度与章质量的估测结合在统的模型中解决。我们通过L-BFGS算法对模型进学习，从获得各个因的权重，并在测试集上对章质量与作者的信誉度进估测以验证模型的正确性，我们选取了个模型作为基准法与我们的法进较，实验显我们的法在准确率和综合评价指标F1上都有较提。在亚马逊评论中，我们提出了评论因图模型，来检测评论中的虚假评论者与虚假评论，该模型将检测虚假评论者和虚假评论结合在统的模型中解决。我们通过L-BFGS算法对模型进学习，获得各个因的权重，并在测试集中检测虚假评论与虚假评论者以验证模型的正确性，我们还选取了个模型作为基准法与我们的法进较，我们的法在准确率和F1上都有较提。章最后总结了本主要作，并提出今后研究向的重点。
[Abstract]:The arrival of Web2.0 promotes the prosperity and development of Hucheng content. While promoting the dissemination of knowledge and solving the problem of unequal information, it also produces problems such as malicious destruction of his edited content, falsification of false comments and disrespect for copyright uploading of his work. These problems make it necessary to evaluate the reputation and content quality. This paper first describes the data acquisition, preprocessing and feature extraction. We take Wikipedia as the representative of knowledge sharing site and download the data of Wikipedia. According to the characteristics of Wikipedia chapter and author we extract quantitative features. We randomly select three classes as three data sets, take the author of chapter and chapter under the data set as the research object, and annotate the quality of chapter and the reputation of the author, and get the standard data set. We use Amazon Review as the representative of the Business Review Station. The research, download Amazon's comment data, and randomly select the comments and reviewers under the electrical products as the research object, first preprocess the data set to solve the problem of data sparsity, according to the characteristics of Amazon comments and reviewers, The quantitative features are extracted and the standard data set is obtained by annotating comments and commenters. In Wikipedia, we propose a biplane factor-graph model, which combines the author's reputation and chapter quality estimation in a unified model. We use the L-BFGS algorithm to study the model, get the weight of each factor, and estimate the chapter quality and the author's reputation on the test set to verify the correctness of the model. We select a model as the benchmark method to compare with our method. The experiment shows that our method is better in accuracy and comprehensive evaluation index F _ 1. In Amazon comments, we propose a review-cause-graph model to detect false reviewers and false comments, which combines the detection of false reviewers and false comments in a unified model. We use the L-BFGS algorithm to learn the model, get the weight of each factor, and test the false comments and the false commenters in the test set to verify the correctness of the model. We also choose a model as the benchmark method to compare with our method. Our method is more accurate and F1. At the end of the chapter, the main works are summarized, and the key points of the future research are put forward.
【学位授予单位】：清华大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP393.09;TP391.1

【共引文献】