Self-training and Label Propagation for Semi-supervised Classification

Abstract

Due to the high cost of manually labeling data and sometimes requiring domain expertise, semi-supervised methods have received a lot of attention. Self-training is a very effective semi-supervised method that greatly improves the problem of insufficient labeled data in classification tasks. In this paper, we propose a semi-supervised classification algorithm based on self-training and label propagation. Specifically, our self-training architecture uses two soft pseudo-labels obtained by the fine-tuned model and label propagation as input to obtain the output of the pseudo-label prediction model, and then selects the high-confidence output of the pseudo-label prediction model as the pseudo-label data. Additionally, we use ImageNet pre-train models for fine-tuning, which greatly reduces learning time and improves accuracy. Experiments show that our method can achieve effective accuracy improvement on a large amount of unlabeled data.