Robust supervised classification algorithm for multivariate contaminated data based on modified M-estimates

Presents a robust Gaussian maximum likelihood (ML) classifier that makes use of the modified M-estimates of the first and second order statistics. The optimal solutions that minimize the modified Huber criterion replace the least squares (LS) estimates. A heuristic approach to deciding the threshold value that influences the estimates and overall classification accuracy is presented. The performance of the robust ML classifier is examined along with the conventional ML classifier for two common contamination cases