An ensemble method for the credibility assessment of user-generated content

Abstract

The Social Web supports and fosters social interactions by means of different social media, which allow the spread of the so called User-Generated Content (UGC). In this context, characterized by the absence of trusted third parties that verify the reliability of the sources and the believability of the content generated, the issue of assessing the credibility of the information diffused by means of social media is receiving increasing attention. In the literature, this issue has been mainly tackled as a classification problem; information is categorized into genuine and fake, usually by implementing or applying classifiers that consider multiple kinds of features (mainly textual and non-textual) to be evaluated in terms of credibility. In this article, unlike prior research, textual features are considered separately with respect to other kinds of features during the classification process. In particular, an Ensemble Method that combines the results produced by two text classifiers and the ones returned by another classifier acting on non-textual features is proposed. This allows to have better results with respect to the use of a single classifier on multiple features together. The effectiveness of the Ensemble Method has been assessed in the context of review sites, by means of a labeled dataset gathered from the Yelp.com site, where on-line reviews are already classified as recommended and not recommended.