Classification performance of mathematical programming techniques in discriminant analysis: Results for small and medium sample sizes

摘要: The performance on small and medium-size samples of several techniques to solve the classification problem in discriminant analysis is investigated. considered are two widely used parametric statistical (Fisher's linear function Smith's quadratic function), a class recently proposed nonparametric estimation based mathematical programming (linear mixed-integer programming). A simulation study performed, analyzing relative above two-group case, for various sample sizes, moderate group overlap across six different data conditions. Training as well validation assess classificatory techniques. degree sizes selected this paper interest practice because they closely reflect conditions many real sets. results experiment show that nonlinear tends be superior training when variances–covariances groups heterogeneous, while technique performs best equal, with equal variances discrete uniform independent variables. also found more sensitive than other size, giving disproportionally inaccurate samples.