This work proposes an original approach using mutual information and Pareto curve jointly for feature selection. Mutual information is used to estimate dependency criterion between features and classes and redundancy criterion between features taken two by two. Unlike some studies, these criteria are used simultaneously to compute Pareto curve and determine the optimal feature sets. This approach is tested on more reference data. Several clustering algorithms are used to compute classification accuracy. The obtained results show the importance of our tools and its ability to select the best feature sets that give the better description.