A Co-training Algorithm by Predicting Confidence of Unlabeled Neighbors

Abstract

As a prominent semi-supervised learning algorithm, co-training can make full use of a few labeled samples as well as a quantity of unlabeled samples to train robust classifiers, thus it is widely researched in the last decades. However, there are still two severe problems in existing co-training methods: firstly, data used in co-training needs at least two sufficient and redundant views, but few practical datasets can meet it. Secondly, because there are a few labeled samples in the initial stage, the initial classifiers trained by co-training are usually too weak to correctly label unlabeled samples, which may bring in noisy labels for the following training process. To deal with these problems, we propose a co-training algorithm by predicting confidence of unlabeled neighbors (termed LCN-CoTrain). Specifically, LCN-CoTrain first defines an entropy-based view division method to generate redundant views. Meanwhile, LCN-CoTrain introduces a novel labeling confidence prediction strategy, which takes the nearest unlabeled neighbors into consideration when calculating the labeling confidence of a certain sample. To validate the efficiency of our proposed LCN-CoTrain algorithm, we experiment on four UCI datasets and compare LCN-CoTrain with several representative co-training methods. The experimental results indicate that our proposed LCN-CoTrain can learn robust classifiers and outperform most of baseline methods.