Robust Deep Graph Based Learning for Binary Classification

Convolutional neural network (CNN)-based feature learning has become the state-of-the-art for many applications since, given sufficient training data, CNN can significantly outperform traditional methods various classification tasks. However, is more challenging if labels are noisy as tends to overfit labels, resulting in sub-par performance. In this paper, we propose a robust binary classifier by CNN-based deep metric functions, construct graph, used clean via graph Laplacian regularization (GLR). The denoised then two proposed loss correction functions regularize functions. As result, node-to-node correlations better reflected, leading improved predictive experiments on three datasets, varying number and type of features under different levels noise, demonstrate that dataset semi-supervised task, our networks several classifiers, including label-noise support vector machine, CNNs with model-based GLR, dynamic classifiers.