Abstract: Tri-Training algorithm belongs to semi-supervised algorithm, the initial classifier performance of the algorithm is influenced by the labeled samples. When the number of the labeled samples is insufficient, the classifier performance is relatively weak, which will affect the iteration later. Therefore, an improvement algorithm called IFS-Tri-Training(Tri-Training based on intuitionistic fuzzy sets) is proposed, the SOM algorithm is applied in the paper to build intuitionistic fuzzy sets, so the classifier can measure the unlabeled samples by more factors, which makes the efficiency of the unlabeled samples is increased and labeled sample set is expanded. The experimental result on UCI datasets shows that the classifier performance is improved and the method of studying unlabeled samples is important.