Structure regularized self-paced learning for robust semi-supervised pattern classification

Nannan Gu, Pengying Fan, Mingyu Fan, Di WangNeural Computing and Applications 2019 · OpenReview

Abstract

Semi-supervised classification is a hot topic in pattern recognition and machine learning. However, in presence of heavy noise and outliers, the unlabeled training data could be very challenging or even misleading for the semi-supervised classifier. In this paper, we propose a novel structure regularized self-paced learning method for semi-supervised classification problems, which can efficiently learn partially labeled training data sequentially from the simple to the complex ones. The proposed formulation consists of three components: a cost function defined by a mixture of losses, a functional complexity regularizer, and a self-paced regularizer; and the corresponding optimization algorithm involves three iterative steps: classifier updating, sample importance calculating, and pseudo-labeling. In the proposed method, the cost function for classifier updating and sample importance calculating is defined as a combination of the label fitting loss and manifold smoothness loss. Then, the importance of the pseudo-labeled and unlabeled samples is adaptively calculated by the novel cost. Unlabeled samples with high importance values are pseudo-labeled with their current predictions. In this way, labels are efficiently propagated from the labeled samples to the unlabeled ones in the robust self-paced manner. Experimental results on several benchmark data sets are provided to show the effectiveness of the proposed method.
BibTeX
@article{gu2019structure,
  author = {Nannan Gu and Pengying Fan and Mingyu Fan and Di Wang},
  title = {{Structure regularized self-paced learning for robust semi-supervised pattern classification}},
  journal = {Neural Computing and Applications},
  year = {2019},
  url = {https://openreview.net/forum?id=yc9MHcEB56},
}

Loading related content...

Figure