I nductive learning is a quickly growing area of machine learning, and the methodology of concept formation is to realize the task of inductive learning incrementally and without supervision. We have proposed a concept formation system CAFE, that creates a concept hierarchy from structured instances. The system uses a new classification algorithm mutual induction and an evaluation function concept-predictability. Our theory is that a system that deals with structures can not only expand possible domains of learning but also use structure information of instances as a domain theory, enabling accurate and fast learning. Such a system can even learn to neglect attributes that do not affect classification, provided appropriate structure information is given. We have evaluated CAFE in two domains (artificial and natural), and the results have proved our theory.