We investigate the problem of item recommendation during the first months of the collaborative tagging community CiteULike. CiteULike is a so-called folksonomy where users have the possibility to organize publications through annotations -tags. Making reliable recommendations during the initial phase of a folksonomy is a difficult task, since information about user preferences is meager. In order to improve recommendation results during this cold start period, we present a probabilistic approach to item recommendation. Our model extends previously proposed models such as probabilistic latent semantic analysis (PLSA) by merging both user-item as well as item-tag observations into a unified representation. We find that bringing tags into play reduces the risk of overfitting and increases overall recommendation quality. Experiments show that our approach outperforms other types of recommenders.