Decision tree (DT) induction is regarded as a representative of traditional approaches to classification rule mining which is an important technique for many data mining applications. Using a heuristic-based local serach, DT induction appends attribute at a time to rules in the order of goodness. This method may eliminate some typical structures that several attributes collectively determine the class. Recently, there has been growing interest in the problem of discovering functional dependencies (FDs) from existing databases [Database Dependency Discovery: a Machine Learning Approach, Tane: an efficient algorithm for discovering functional and approximate dependencies, Efficient Discovery of Functional Dependencies and Armstrong Relations, an efficient algorithm for mining functional and embedded dependencies]. Some efficient and scalable algorithms have been proposed. In this paper, we present a new method to build a DT classifier using approximate FDs [Tane: an efficient algorithm for discovering functional and approximate dependencies]. The new method is different from the traditional ways of building DTs in that it searches composite attributes for individual node of a DT which leads to substantially smaller and more understandable DTs without adversely affecting the accuracy gains. Experiments showed that the new method not only builds more accurate classifiers, but also does this with more compact structures.