Many scientific fields analyzing and modeling social networks have focused on manually-collected datasets where the friendship links are sparse (due to the costs of collection) but relatively noise-free (i.e. they indicate strong relationships). In online social networks, where the notion of ``friendship'' is broader than what would generally be considered in sociological studies, the friendship links are denser but the links contain noisier information (i.e., some weaker relationships). However, the networks also contain additional transactional events among entities (e.g., communication, file transfers) that can be used to infer the true underlying social network. With this aim in mind, we develop a supervised learning approach to predict link strength from transactional information. We formulate this as a link prediction task and compare the utility of attribute-based, topological, and transactional features. We evaluate our approach on public data from the Purdue Facebook network and show that we can accurately predict strong relationships. Moreover, we show that transactional-network features are the most influential features for this task.