Manifolds for training set selection through outlier detection

The effect of the training set on supervised classifier performance has always been overlooked. This paper provides a new approach for training set cleaning based on the concept of outlier detection to help build sound class models during the training of supervised classifiers. Outliers in a training set result in classifier performance deterioration and slow convergence. For training set cleaning, the proposed technique transforms non-linear relationships between high dimensional patterns into a simple geometric relationship. The Isometric pattern Mapping (ISOMAP) is used to embed the high dimensional training set patterns to a low-dimensional manifold. The dispersion of mapped points will be used to locate the outliers and measure their outlyingness. Several experiments on real data sets show the promising performance of the proposed technique.