The study of recommender systems is essential nowadays due to its great effect on businesses and customer satisfaction. Different active learning strategies were previously developed to gain ratings from the users on specific items, and this enables the system to have more information and consequently make more accurate recommendations. In previous studies, these strategies were evaluated using a different selection of metrics in each work, and the experimentations were done on different datasets. In this paper, we solve these weaknesses by comparing the main ten non-personalized strategies on a fair ground, by simulating them against two datasets and using seven of the mostly agreed upon metrics. This gives more trust and less biased results when comparing their performances. Also, the analysis of the computation time and the elicitation efficiency is added.