Negative $\varepsilon$ Dragging Technique for Pattern Classification

In this paper, we propose the negative ε dragging technique for robust classification of noisy and contaminated data. Different from the naïve ε dragging technique, the negative ε dragging technique argues that robust results can be obtained by properly reducing the class margin of conventional least squares regression when performing classification on noisy data. The underlying rationale of the negative ε dragging technique assumes that setting a relative small class margin for the training procedure of least squares regression leads to desirable generalization capability, which, therefore, considerably contributes to boosting the classification performance for the data corrupted with noise. The experimental results indicate that our technique obtains better classification accuracy.