Approaches to Enhancing Semi-Supervised Learning using Process Data Augmentation via Self-Labeling and Generative Adversarial Networks
Prince Addo, Vinay Prasad · IFAC-PapersOnLine · 2024
Augmentation of limited data with additional new data has been of major importance in fields where available data is scarce or expensive to acquire, and this is achieved via either the generation of new synthetic data or the labelling of pre-existing unlabelled data. Data generation using generative adversarial networks (GANs) and their variants have been widely reported to produce useful synthetic data. Self-labeling of unlabelled data has also been successfully implemented to provide labels to augment the training set of limited datasets. In this work, we explore these approaches for improving the new data quality (either generated or self-labeled) for regression and classification problems using various experiments with a specific focus on froth flotation data. Our work proposes a combination of regressor and classifier networks with shared network layers both for generation of new data and self-labeling of unlabeled data. We show that this approach provides superior performance in comparison to just using a classifier network.