Multi-scale fuzzy feature selection method applied to wood singularity identification

Abstract

A multi-scale feature selection method based on the Choquet Integral is presented in this paper. Usually, aggregation decision-making problems are well solved, relying on few decision rules associated to a small number of input parameters. However, many industrial applications require the use of numerous features although not all of them will be relevant. Thus, a new feature selection model is proposed to achieve a suitable set of input features while reducing the complexity of the decision-making problem. First, a new criterion, combining the importance of the parameters as well as their interaction indices is defined to sort them out by increasing impact. Then, this criterion is embedded into a new random parameter space partitioning algorithm. Last, this new feature selection method is applied to an industrial wood singularity identification problem. The experimental study is based on the comparative analysis of the results obtained from the process of selecting parameters in several feature selection methods. The experimental study attests to the relevance of the remaining set of selected parameters.