Chernoff FDA for Heterocedastic Gaussians
FDA does not minimize the classification error when the probability distributions of class samples are heterocedastic Gaussians (see the following case).
FDA classifier
To minimize the Chernoff distance between both distributions, we can minimize the classification error when both distributions are general Gaussians.
See the result of Chernoff FDA :
Their kernel classification model:
- KFDA
- Kernel CFDA