Robust Adaptive Semi-supervised Classification Method based on Dynamic Graph and Self-paced Learning

Despite the computers have developed rapidly in recent years, there are still many difficulties to obtain a large number of labelled data in many practical problems, for example, medical image diagnosis, internet fraud, and pedestrian detection. To deal with learning problems with only a few labeled data, a novel semi-supervised learning method combined with dynamic graph learning with self-paced learning mechanism is present in this work, namely SS-GSELM. Firstly, according to the loss value of labeled samples in each training, the algorithm selects the sample with the smaller loss value for learning, and then gradually adds the sample with the larger loss value during the training process until all labeled samples are trained. In particular, different weights are given to samples through a regularization function to adjust the importance of different samples on the model results. Secondly, the algorithm uses local consistency property as supplementary information to enhance the performance of the learning machine, so an adaptive graph matrix is constructed to retain data similarity information. To do this, an alternative strategy is proposed to update graph matrices and self-paced weights to adapt to the classifier. Experimental results on real data sets exhibit that the proposed method superior to the classic methods in classification tasks.