The multi-label classification got recent interest in the machine learning community by its usefulness in many areas. As with any machine learning problem, the need to preprocess multi- label data has emerged as a need to improve the performance of learners. In this paper, we introduce a new method selection and variable processing for multi-label classification. This method is an adaptation of MDL criterion and is based on a Bayesian approach. A comparative study is made with other methods of the state of the art to position the new method but also to show interest of the features selection for the multi-label classification.