Abstract:

Taking full advantage of  three machine learning methods,  active learning, semi-supervised learning and ensemble learning, an efficient learning algorithm was proposed. The algorithm did  not require two sufficient and redundant views, which had  high generalization abilities. To reduce the error rate, based on the clustering assumption, a new approach was presented to estimate the degree of confidence of the labeled samples in each iteration of the co-training process. Also, a new measure of samples,  the degree of contribution, was given as a clue for selecting the unlabeled samples. Since a high degree of contribution implies a great value of the sample to be selected, selecting the samples with high degrees of contribution after each iteration could  enhance the  feedback effect and the  learning performance. All these could lead to a new ensemble co-training algorithm based on active learning. The feasibility and the performance of the algorithm  were verified by image retrieval experiment.

Key words: co-training, semi-supervised learning, ensemble learning, active learning, image retrieval