A Study on Deep Semi-supervised learning method using Data-adaptive Augmentation Technique

Abstract

This study introduces a novel semi-supervised learning approach specifically designed for tabular data, featuring a unique learnable data augmentation technique that preserves the labeled data’ information. The approach is mainly motivated by two methods: MixMatch, known as one of the state-of-the-art semi-supervised learning methods in image data, and Neutral AD, a self-supervised learning method for anomaly detection. These inspirations are adapted to tabular data through an innovative loss function comprising three distinct parts: one for labeled data, one for unlabeled data, and another for deterministic contrastive learning. This loss function is pivotal in guiding transformations that produce diverse and informative data augmentations, while preserving the characteristics of the original data. To validate our proposed method, we perform experiments on three tabular datasets, where our method demonstrates remarkable state-of-the-art performance, especially on the two datasets. The results not only show superior test accuracy over several baselines, but also highlight the importance of each components’ role by the tuning hyperparameters proposed in the ablation studies.