The problem of social spam detection has been traditionally modeled as a supervised classification problem. Despite the initial success of this detection approach, later analysis of proposed systems and detection features has shown that, like email spam, the dynamic and adversarial nature of social spam makes the performance achieved by supervised systems hard to maintain. In this paper, we investigate the possibility of using the output of previously proposed supervised classification systems as a tool for spammers discovery. The hypothesis is that these systems are still highly capable of detecting spammers reliably even when their recall is far from perfect. We then propose to use the output of these classifiers as prior beliefs in a probabilistic graphical model framework. This framework allows beliefs to be propagated to similar social accounts. Basing similarity on a who-connects-to-whom network has been empirically critiqued in recent literature and we propose here an alternative definition based on a bipartite users-content interaction graph. For evaluation, we build a Markov Random Field on a graph of similar users and compute prior beliefs using a selection of state-of- the-art classifiers. We apply Loopy Belief Propagation to obtain posterior predictions on users. The proposed system is evaluated on a recent Twitter dataset that we collected and manually labeled. Classification results show a significant increase in recall and a maintained precision. This validates that formulating the detection problem with an undirected graphical model framework permits to restore the deteriorated performances of previously proposed statistical classifiers and to effectively mitigate the effect of spam evolution.