Feature space reduction and parameter optimization with application to semi-supervised water pollutant classification

This paper focuses on optimizing the classifier and the used feature space of a semi-automatic approach for water floating agents classification. Our approach proposes a method for pollutant/non pollutant river waste labeling. For this task we consider a soft-margin kernel SVM classifier. The optimization process consists in determining the optimal features for classification with regard to two scores: accuracy and recall of the classifier. Then, a series of experiments are considered in order to decide on the most effective kernel that transforms our data space into a linear one. Finally, experiments for optimally tuning the classifier parameters are presented.