Semi-supervised context characterized by the presence of a few pairs of constraints between learning samples is abundant in many real applications. Analysing these instance constraints by recent spectral scores has shown good performances for semi-supervised feature selection. The performance evaluation of these scores is generally based on classification accuracy and is performed in a ground truth context. However, this supervised context used in the evaluation is inconsistent with the semi-supervised context used in the feature selection. In this paper, we propose a semi-supervised performance evaluation procedure, so that both feature selection and clustering take into account the constraints given by the user. In this way, the selection and the evaluation steps are performed in the same context which is close to real life applications. Extensive experiments on benchmark datasets are carried out in the last section. These experiments are performed using a supervised classical evaluation and the semi-supervised proposed one. They demonstrate the effectiveness of feature selection based on constraint analysis that uses both pairwise constraints and the information brought by the unlabeled data.