Abstract:

The current implementation of KLDA in R fails to compute projections in the case that the kernel matrix is noninvertible in the objective function. We propose an algorithm for adjusted KLDA which allows for the approximation of singular matrices within KLDA’s objective function, ensuring the success of computations for any set of tuning parameters.The validity of the algorithm is evaluated on several simulated datasets, then applied to three versions of a subset of the MorphII dataset containing different extracted features for face imaging tasks. The transformed feature set is used to train several statistical classification models, whose performance is then evaluated to determine the efficacy of the algorithm.
