Facebook Profile Credibility Detection using Machine and Deep Learning Techniques based on User’s Sentiment Response on Status Message

Abstract

Recently, the impact of online Social Networks sites (SNS) has dramatically changed, and fake accounts became a vital issue that has rapidly evolved. This issue gives rise to how to assess and measure the credibility of User-Generated Content (UGC). This content is used in finding trusted sources of information on SNS like Facebook, Twitter, etc. Consequently, classifying users’ profiles and analyzing each user’s behavior response based on the content generated became a challenge that must be solved. One of the most significant approaches is Sentiment Analysis (SA) which plays a major role in assessing and detecting the credibility degree of each user account behavior. In this paper, the aim of the study is to measure and predict the user’s profile credibility by declaring the correlation degree among the UGC features that affect users’ responses to status messages. The proposed models were implemented using six Supervised Machine Learning classifiers, an Unsupervised Machine Learning cluster model, and a Deep Learning Neural Network (NN) model. The research paper presents two experiments to evaluate Facebook profile credibility. At first, we applied a binary classification model to classify profiles into fake or genuine users. Then, we conducted a classification model on genuine users based on the credibility theory by using the Analytical Hierarchical Process (AHP) approach and computed the credibility score for each. Secondly, we selected and analyzed a public Facebook page (CNN public page) and obtained data from it for users’ sentiment reactions and responses on statuses Messages relating to different topics on the period (2016/2017). Then, we performed LDA on the status corpus (Topic Modeling algorithm, Latent Dirichlet Allocation) to generate topic vectors. In addition, we performed Principal Component Analysis (PCA) method to visualize and classify each status topic distribution. Afterthought, we produced a status corpus cluster to classify users’ behaviors through statuses posted and users’ comments. As a conclusion of this study, the first experimental results achieved 95% and 99% accuracy to classify fake/genuine users and incredible/credible accounts, respectively. The second experiment outcome identified the clusters for the status corpus in 10 topic-features distribution and classified users’ contents into credible or not according to the final calculated credibility score.