This article proposes an evidential combination method with the learned optimal weights to take full advantage of the knowledge in the labeled and unlabeled data. The new method can reduce the negative influence of high-confidence patterns with wrong predictions. It is a co-training-like paradigm to solve the classification problem with few labeled and massive unlabeled patterns. Moreover, the combination of classification results and clustering results is not considered in many previous works. The effectiveness of CILU was evaluated with respect to a variety of advanced methods, and the experiments show CILU can significantly improve the classification accuracy. This framework will inspire many researchers to develop some new algorithms for semi-supervised learning.

For classification with few labeled and massive unlabeled patterns, co-training, which uses information in labeled and unlabeled data to classify query patterns, is often employed to train classifiers in two distinct views. The classifiers teach each other by adding high-confidence unlabeled patterns to training dataset of the other view. Whereas, the direct adding often leads to some negative influence when retraining classifiers because some patterns with wrong predictions are added into training dataset. The wrong predictions must be considered for performance improvement. To this end, we present a method called Combination of Information in Labeled and Unlabeled (CILU) data based on evidence theory to effectively extract and fuse complementary knowledge in labeled and unlabeled data. In CILU, patterns are characterized by two distinct views, and the unlabeled patterns with high-confidence predictions are first added into the other view. We can train two classifiers by few labeled t...