Exploring Latent Sparse Graph for Large-Scale Semi-Supervised Learning

Abstract

We focus on developing a novel scalable graph-based semi-supervised learning (SSL) method for input data consisting of a small amount of labeled data and a large amount of unlabeled data. Due to the lack of labeled data and the availability of large-scale unlabeled data, existing SSL methods usually either encounter suboptimal performance because of an improper graph constructed from input data or are impractical due to the high-computational complexity of solving large-scale optimization problems. In this paper, we propose to address both problems by constructing a novel graph of input data for graph-based SSL methods. A density-based approach is proposed to learn a latent graph from input data. Based on the latent graph, a novel graph construction approach is proposed to construct the graph of input data by an efficient formula. With this formula, two transductive graph-based SSL methods are devised with the computational complexity linear in the number of input data points. Extensive experiments on synthetic data and real datasets demonstrate that the proposed methods not only are scalable for large-scale data, but also achieve good classification performance, especially for an extremely small number of labeled data.