Complex objects are organized into class/subclass hierarchy where each object attribute may be composed of other complex objects. Almost of the existing works on complex data classification start by generalizing objects in appropriate abstraction level before the classification process. Generalization prior to classification produces less accurate result than integrating generalization into the classification process. This paper proposes CO4.5, an approach for generating decision trees for complex objects. CO4.5 classifies complex objects directly through the use of inheritance and composition relationships stored in object-oriented databases. Experimental results, using large complex datasets, showed that CO4.5 yielded better accuracy compared to traditional data classification techniques.