We present an active meta learning approach to model selection or algorithm recommendation. We adopt the point of view "collab-orative filtering" recommender systems in which the problem is brought back to a missing data problem: given a sparsely populated matrix of performances of algorithms on given tasks, predict missing performances; more particularly, predict which algorithm will perform best on a new dataset (empty row). In this work, we propose and study an active learning version of the recommender algorithm CofiRank algorithm and compare it with baseline methods. Our benchmark involves three real-world datasets (from StatLog, OpenML, and AutoML) and artificial data. Our results indicate that CofiRank rapidly finds well performing algorithms on new datasets at reasonable computational cost.