Adaptive Graph Learning for Semi-supervised Self-paced Classification

Semi-supervised learning techniques have been attracting increasing interests in many machine learning fields for its effectiveness in using labeled and unlabeled samples. however, the ultimate performance tend to be inaccurate or misleading due to the presence of heavy noise and outliers. This problem raises the need to develop the methods that can exploit data structure and also be robust to the noisy points. In this paper, a novel semi-supervised classification method, named adaptive graph learning for semi-supervised self-paced classification (AGLSSC in short), is proposed by integrating self-paced learning (SSL) regime and adaptive graph learning (AGL) strategy into a joint framework and experimentally evaluated. Specifically, AGLSSC automatically select import samples by adding a parameter that can measure the importance of samples in each iteration optimization process. In addition, in order to learn the internal relationship of samples from corrupt data, the proposed method adaptively learns an optimal sample similarity matrix while maintaining the local structure of the samples. In this case, the proposed model has strong robustness to noise points. Extensive experiments conducted on diverse benchmarks demonstrate that AGLSSC achieves the most outstanding performance compared to some state-of-the-art semi-supervised classification methods.