CoPASample: A Heuristics Based Covariance Preserving Data Augmentation

Rishabh Agrawal, Paridhi Kothari

Abstract: An efficient data augmentation algorithm generates samples that improves accuracy and robustness of training models. Augmentation with informative samples imparts meaning to the augmented data set. In this paper, we propose CoPASample (Covariance Preserving Algorithm for generating Samples), a data augmentation algorithm that generates samples which reflects the first and second order statistical information of the data set, thereby augmenting the data set in a manner that preserves the total covariance of the data. To address the issue of exponential computations in the generation of points for augmentation, we formulate an optimisation problem motivated by the approach used in ν-SVR to iteratively compute a heuristics based optimal set of points for augmentation in polynomial time. Experimental results for several data sets and comparisons with other data augmentation algorithms validate the potential of our proposed algorithm.