The accuracy of recommendations of collaborative filtering bas\-ed recommender systems mainly depends on which users (the neighbors) are exploited to estimate a user's ratings. We propose a new approach of neighbor selection, which adopts a global point of view. This approach defines a unique set of possible neighbors, shared by all users, referred to as Global Neighbors ($GN$). We view the problem of defining $GN$ as a combinatorial optimization problem and propose to use an evolutionary algorithm to tackle this search. Our aim is to find a relatively small $GN$ as the size of the resulting model, as well as the complexity of the computation of recommendations highly depend on the size of $GN$. We present experiments and results on a standard benchmark data-set from the recommender system community that support our choice of the evolutionary approach and show that it leads to a high accuracy of recommendations and a high coverage, while dramatically reducing the size of the model (by 84\%). We also show that the evolutionary approach produces results able to generate accurate recommendations to unseen users, while easily allowing the insertion of new users in the system with little overhead.