Online Social networks have gained much prominence in the recent years such that it has become an unavoidable means of daily communication. The element of trust in social networks has been studied ever since the inception of online social networks. Trust in online social networks is extremely fragile in nature due to the virtual connections between users in the network. The level of trustworthiness of each user in a social network varies and is usually computed using reputation level of the users. This paper focuses on identifying the features that determine the trust of a user in online social networks using benchmark datasets. We propose a new probabilistic reputation feature model that is better than the raw reputation features. The enhanced trust prediction framework has been tested and validated on three benchmark datasets namely Wikipedia election dataset, Epinions dataset and Slashdot dataset. The proposed probabilistic feature enhances the overall accuracy, F1 score, and area under the ROC for the classifier results significantly. The results have been compared with other state of the art techniques and are found to be efficient. © 2016 Elsevier B.V.