Emotions can be understood as automatic responses from different biological systems of the human body to a particular internal or external stimulus. From different studies, there have been proposed emotions classification schemes as the discrete space, where every emotion is assigned a particular tag and dimensional spaces in which an emotion is determined as a combination of two indices which are related to a more accurate measurement and quantification of emotions. Based on a given classification space there has been developed different research projects in order to gave a machine the ability to detect and recognize the emotions of humans, considering that understand and differentiate the emotions expressed by people is still a complex task even for human beings themselves. Because emotions produce different biological responses of the human body, it is intuitive to think of the development of systems that incorporate biological information from these systems to make effective recognition of emotions. The multimodal approach for emotion recognition then comprises those methodologies that combine characteristics of different signals obtained from the human body to a given stimulus that causes a specific emotional state. The signals that are commonly used in the development of multimodal systems are the video information, the audio signal, the electroencephalogram (EEG) and various physiological signals such as heart rate, galvanic skin response , the respiratory signal and the temperature. In this paper we propose a feature selection methodology within a multimodal approach for the classification of emotional states in a dimensional space, based on discriminative models by implementing two selection algorithms known as Recursive Feature Elimination (RFE) and margin-maximizing feature elimination (MFE). With these approaches for feature reduce the original set of features is reduced from the analysis of the linearity and nonlinearity of the different signals. Nonlinear analysis is based on certain metrics obtained from a technique known as recurrence plots (RP) that have showed to improve the classification results in some state of art works. We use two databases recognized within the state of the art for the extraction and selection of features within the proposed scheme.