Discrete kernels, loss functions and parametric models in discrete discrimination: A comparative study

Two kernel-based approaches to discriminant analysis are considered: the traditional one where kernels are used to estimate the distribution of the predictor variables given the group and a direct kernel method where kernels are used to estimate the a posteriori probabilities directly. For both approaches cross-validatory choice of smoothing parameters is based on various loss functions which are directly connected to the separation of groups. Comparison with parametric models shows the improvement gained by the more flexible kernel approaches.