Problem of classification is the main research target of many algorithms in machine learning and data mining. Of all the algorithms, decision tree is more preferred by researchers due to its clarity and readability. Attribute of little value domain is the important feature of training dataset of decision trees. Based on this, this paper presents a new approach to construct decision tree after reducing dimension and compressing data set. Experiment shows that the algorithm proposed in this paper improves the efficiency in real applications compared with traditional algorithms.