Semi-supervised Learning for Multi-label Classification

In this report we consider the semi-supervised learning problem for multi-label image classification, aiming at effectively taking advantage of both labeled and unlabeled training data in the training process. In particular, we implement and analyze various semi-supervised learning approaches including a support vector machine (SVM) method facilitated by principal component analysis (PCA), and a self-training method that iteratively conducts supervised learning and enlarges the set of training labels on the go. We compare the performances of semi-supervised learning methods with supervised learning benchmarks, and introduce a heuristic performance analysis for the training process. In addition, we analyze the impact of different training parameters for the PCA-SVM and the self-training method on the prediction performance. The algorithms are implemented for the ChestX-ray14 [32] medical image dataset.