Features influencing the concept of trust in online reviews

The number of online reviews increase considerably on platforms and have a significant impact on purchase decisions. These reviews can represent both an opportunity and a threat for a company. It is therefore essential to detect among the huge quantity of reviews, those which are unreliable. A question then arises: how to detect deceptive reviews? In this paper, we were interested in the concept of review trustworthiness. To detect the reliability of reviews, we give our definition of the reliability and propose a machine learning approach based on different classifiers. We consider that the information related to the comment itself (sentiments, linguistic elements etc.), and that relating to the user (activity, experience, sociability etc.) play a role on reliability of the review. We propose to include in our method these two classes of information. We did a series of experiments on the Yelp open dataset. Furthermore, in these experiments we have compared the performances of our trust model with a second model based only on the review's content (using TF-IDF vectorisation method). Our results show on the one hand that the information we used (comments+users) play a role in determining the reliability of the review, and on the other hand that we achieve better performances with our features (i.e. without using TF-IDF method).