Efficient Data Augmentation for Machine Learning Classifiers Using Horizontal and Vertical Contractions

To yield better performance, machine learning (ML) classifiers are typically trained with augmented data (i.e., combining real data with synthetically generated data). While the amalgamation of real data with synthetic data makes intuitive sense and can be useful in enhancing the performance of ML classifiers, the addition of synthetic records beyond a certain limit destroys the truthfulness of the real data and can lead to performance bottlenecks. To address these problems, we propose and implement an efficient data augmentation technique based on vertical and horizontal contractions, which adds as few records as possible (under salient features only), whereas existing methods augment everything, leading to high computation costs. Our technique identifies suitable features and regions for augmentation without losing guarantees on performance, rather than needlessly adding more records. Experiments with three real-world datasets prove the efficacy of our technique in terms of computing time, accuracy, and model complexity.