Face recognition systems aimed at working on large scale datasets are required to solve specific hurdles. In particular, due to the huge amount of data, it becomes mandatory to furnish a very fast and effective approach. Moreover the solution should be scalable, that is it should deal efficiently the growing of the gallery with new subjects. In literature, most of the works tackling this problem are composed of two stages, namely the selection and the classification. The former is aimed at significantly pruning the face image gallery, while the latter, often expensive but precise, determines the probe identity on this reduced domain. In this article a new selection method is presented, combining a multi-feature representation and the least squares method. Data are split into sub-galleries so as to make the system more efficient and scalable. Experiments on the union of four challenging datasets and comparisons with the state-of-the-art prove the effectiveness of our method.