Linear models for cost-sensitive classification

摘要: In this paper, we investigate the performance of statistical, mathematical programming and heuristic linear models for cost-sensitive classification. particular, use five techniques including Fisher's discriminant analysis DA, asymmetric misclassification cost mixed integer AMC-MIP, support vector machine CS-SVM, a hybrid SVMIP genetic algorithm CGA techniques. Using simulated datasets varying group overlaps, data distributions class biases, real-world from financial medical domains, compare performances our based on overall holdout sample cost. The results experiments indicate that when overlap is low distribution exponential, DA appears to provide superior performance. For all other situations with datasets, CS-SVM provides case domain, AMC-MIP hold slight edge over two SVM-based classifiers. However, domains continuous discrete attributes, SVM classifiers perform better than model most computationally inefficient poor performing model.