This paper presents a feature selection comparison oriented to human action recognition only with the kinematic features of skeleton representation. For this purpose, three relevance methods are used to rank the contribution of kinematic features for classifying an action is employed. Particularly, the method with the best results includes the supervised information regarding the action to find out a relevant set of features, encoding the most discriminative information. Experimental results are obtained on a well-known public data (MSR Action3D). Results are encouraging to use kernel theory methods to get better kinematic feature selection for each action with a good generalization indistinct to the subject