Com binin g Classifier Gui ded by Sem i-Su pervis ion

The articlee suggestss an algorithm for regular c lassifier ensemble methodology. The proposed methodol ogy is based on po ssibilistic aggregation to cla ssify samples. The argued m ethod op timizes an objective function that combines environment recognition, multi-criter ia aggregation ter m and a learning term. The optimization aims at learnin g backgrounds as solid clussters in subspaces of the high-dimensiional feature-space via an u nsupervised learning includding an a ttribute discrimin ation component. The unsupervised cl ustering component assigns degree o f typicality to each data pattern in or der to identify and reduce thhe effect of noisy or outlaid data patterns. Then, the suggested techhnique obtains the best combination p arameters for each background. The experimentations on artificia l datasets and sta ndard SONAR dataset demonnstrate thhat our classifier ensemble does better than individual classifiers in th e ensemb le.