Robust Graph Learning for Semi-Supervised Classification

Learning graphs from data automatically has shown impressive performance on semi-supervised classification. However, the real data is often corrupted, which can lead to the learned graphs being inexact or unreliable. This paper proposes a robust graph learning scheme by removing noise and errors in the raw data adaptively. The proposed model can also be seen as an adaptive manifold regularized robust PCA with the quality of graph playing a pivotal role. It can boost the performance of semi-supervised classification due to two key factors: 1) enhanced low-rank recovery by using the graph smoothness assumption, 2) improved graph construction by using clean data recovered by robust PCA. Extensive experiments on semi-supervised classification show that our model outperforms other state-of-the-art algorithms.