For the performance degradation of unlabeled data in semi-supervised learning, a new cooperative training
algorithm, LDL-tri-training, is proposed. Firstly, by using least significant difference(LSD) hypothesis testing method,
significant differences among three classifiers are tested. Then a D-S evidence theory is adopted to improve the stability
of unlabeled data. Finally, local outlier factor(LOF) algorithm is used to reject error labeled data. Experiments show that
LDL-tri-training can more effectively and stably utilize the unlabeled examples to improve classification generalization.