The rapid growth of data volumes shared on location-based social networks (LBSN) enables the extraction of users' preferences. Then those preferences can be used to rec- ommend to the user a list of points-of-interest matching his profile. Today the recom- mendation of points-of-interest has become an essential component of LBSN. Unfor- tunately traditional recommendation methods fail to adapt to the specific constraints of LBSN such as the high sparsity of the data, or to take into account the geographical influence. In this paper we present a model of recommendation based on the Poisson factorization that offers an effective solution to these constraints. We have tested our model through experiments on a realistic data set from the LBSN Foursquare. These experiences have enabled us to demonstrate a better recommendation than 3 models of state-of-the art.