Semi-supervised Co-training Algorithm Based on Assisted Learning

Abstract

The classification performance of the learner is weakened when unlabeled examples are mislabeled during co-training process. A semi-supervised co-training algorithm based on assisted learning (AR-Tri-training) was proposed. Firstly, the assisted learning strategy was presented, which is combined with rich information strategy for designing the assisted learner. Secondly, the evaluation factor was calculated, and noise was eliminated from unlabeled example set by using the assisted learner and the evaluation factor. Finally, three single learners were trained using labeled examples, wrong-learning examples on validation set and less noise unlabeled examples. The experimental results on application to voice recognition indicate that AR-Tri-training can compensate for the Tri-training shortcomings and the average classification accuracy is increased by 15%. As can be drawn from the experimental results, AR-Tri-training not only removes the mislabeled examples in training process, but also takes full advantage of the unlabeled examples and wrong-learning examples on validation set.