Social spam has been plaguing online social networks for years. Being the sites where online users spend most of their time, the battle to capitalize and monetize users' attention is actively fought by both spammers and legitimate sites operators. Social spam detection systems have been proposed as early as 2010. They commonly exploit users' content and behavioral characteristics to build supervised classifiers. Yet spam is an evolving concept, and developed supervised classifiers often become obsolete with the spam community continuously trying to evade detection. In this paper, we use similarity between users to correct evasion-induced errors in the predictions of spam filters. Specifically, we link similar accounts based on their shared applications and build a Markov Random Field model on top of the resulting similarity graph. We use this graphical model in conjunction with traditional supervised classifiers and test the proposed model on a dataset that we recently collected from Twitter. Results show that the proposed model improves the accuracy of classical classifiers by increasing both the precision and the recall of state-of-the-art systems.