Avaliando atributos de credibilidade de páginas Web utilizando Aprendizagem de Máquina

Abstract

Information shared on the Web propagates quickly, whether true or not. Credibility in this context refers to the level of trust a user places subjectively on a Web page. The purpose of reproducing incorrect information is related to several factors such as political manipulation, obtain financial benefits, disseminate malicious defamation, among others. Therefore, verifying the credibility of the information available on the Web ends up being a mandatory task. Among the various techniques developed to detect whether aWeb page can be accredited or not, machine learning is the most used in comparison to the assessment of credibility manually. The purpose of this work is to evaluate and define attributes that can be used in a future model for assessing the credibility of Web pages, by extracting characteristics from the content of the page and the network, with the help of machine learning classifiers, thus enabling greater certainty on the credibility of web pages. As a result, this dissertation concluded that the Random Forest classifier had the best result for assessing the credibility of web pages with 95.36% accuracy. In addition to providing an attribute extraction script, also pointing out which are the most relevant and easy extraction attributes that can be selected for any URL, for that, 3 attribute selection methods are used: Select the best, RFE Selection and Selection RFECV, the last result with 95.33% accuracy.