Probabilistic Multimodal Classification with dynamic feature selection

A novel Probabilistic,Multimodal Classification with Dynamic Input Selection (PMC-D) framework is proposed in this paper. In this approach, Probability Distribution Functions (PDF), which are determined for each i) modality´s input feature and ii) output class, are used to score per class the importance of each given input feature value. Furthermore, these scores/beliefs, which are dependent on the instantaneous (dynamic) values of input features, are used to reduce classification dimensionality and increase classification performance. PMC-D is generic and does not require weighting or normalization of feature scores. Moreover, using simulated Gaussian and non-Gaussian PDF types of input datasets, as well as datasets provided from three well-known real applications, experimental results have shown that the PMC-D methodology offers classification advantages when compared to well-known classification techniques.