Semi-Supervised Learning with Per-Class Adaptive Confidence Scores for Acoustic Environment Classification with Imbalanced Data

In this paper, we concentrate on the per-class accuracy of neural network-based classification in the context of identifying acoustic environments. Even a fully supervised learning framework with an equal amount of data for each class can lead to significant differences in class accuracies. This is then amplified by semi-supervised learning using naturally imbalanced data. To address this problem, we propose an adaptive method for pseudo-label selection via a straightforward optimization of the validation accuracy per class, aimed specifically at reducing the variance between different classes. The proposed method is general and can be applied for both maximum probability and entropy-based confidence criteria. Compared to fully supervised learning as well as state-of-the-art methods for pseudo-labeling, it achieves the lowest variances of per-class accuracy and the highest accuracies of the minority classes when tested on common publicly available environment sound databases.