A co-training method based on parameter-free and single-step unlabeled data selection strategy with natural neighbors

As an effective semi-supervised learning algorithm, the co-training method trains two classifiers on two views independently. The unlabeled sample selection strategy in the self-labeled process is crucial for co-training. However, most of the existing strategies strongly depend on parameter settings and require re-calculating the confidence of unlabeled samples in each iteration. Inspired by the concept of natural neighbors introduced recently, a co-training method based on parameter-free and single-step unlabeled data selection strategy with natural neighbors (CT-NaN) is proposed in this paper. In CT-NaN, the confidence value of unlabeled samples is calculated in a parameter-free manner by analyzing the training data based on natural neighbors before the iteration of co-training, and it requires to be calculated only once in the whole process of co-training. Besides, CT-NaN is able to mitigate the negative effect of outliers because the training stops automatically when only outliers remain. Four groups of experiments with 22 data sets are conducted, and the results verify the effectiveness of CT-NaN when compared with 8 state-of-the-art co-training methods.