The manual annotation of images remains a major challenge in building robust classifiers, especially in the Big Data era where labeled data are limited. Graph-based semi-supervised learning (SSL) provides an effective solution by leveraging both labeled and unlabeled data. In this paper, we propose a novel graph-based SSL framework that integrates labeled samples with abundant unlabeled data through an adaptive graph structure. The proposed method jointly learns three components: an auto-weighted low-rank graph, soft labels for unlabeled data, and a discriminative latent subspace. By incorporating soft labels into the subspace learning process, the model enforces consistency between the graph structure and the data manifold. This leads to improved discriminative representation. Unlike traditional approaches that treat these components separately, our method optimizes them jointly within a unified framework. Experimental results on multiple benchmark datasets demonstrate the effectiveness of the proposed approach, achieving superior performance under different labeling conditions.
Baradaaji et al. (Mon,) studied this question.