Co-training is a semi-supervised learning method, which employs two complementary learners to label the unlabeleddata for each other and to predict the test sample together. Previous studies show that redundant information can helpimprove the ratio of prediction accuracy between semi-supervised learning methods and supervised learning methods. However,redundant information often practically hurts the performance of learning machines. This paper investigates what redundantfeatures have effect on the semi-supervised learning methods, e.g. co-training, and how to remove the redundant features aswell as the irrelevant features. Here, FESCOT (feature selection for co-training) is proposed to improve the generalizationperformance of co-training with feature selection. Experimental results on artificial and real world data sets show that FESCOThelps to remove irrelevant and redundant features that hurt the performance of the co-training method.