Microbial data augmentation combining feature extraction and transformer network

Microbial data exhibit high dimensionality, feature sparseness, and class imbalance. Popular data augmentation strategies typically generate intersecting or overlapping samples with low diversity. In this paper, we propose a two-stage data augmentation method to synthesize high-quality samples. First, we train a feature extractor by minimizing the cross-entropy. Positive training instances are oversampled to achieve class balance. Second, a transformer network is trained for data augmentation to balance diversity and discernibility. A dropout technique is designed to randomly set some feature values as missing. Experiments were carried out on 10 microbial datasets. The results show that: (1) the constructed feature extraction neural network can effectively reduce the dimension of the data, make the data no longer sparse, improve the distinguishability of the samples, and obtain more obvious classification boundaries, and (2) The trained data augmentation transformer network with dropout technique can generate high quality samples, improve the performance of the classifier and reduce the cost of misclassification.