It is an object of the present invention to find out parts to be a highly possible cause of failure without searching all of part data of all of products.Dispersed parts data on a parts tree are sequentially accessed from a set of known failed products, and part attribute values each having a higher support in the faulty product are extracted. In this process, a subset of parts used in the faulty product is also obtained simultaneously. The part attribute values having higher supports and the subset of parts used in the faulty product are represented as a tree in which a parts type serves as a node. Next, an information gain of a rule that having the two part attribute values is a cause of failure is calculated on two part attribute values having higher supports on the tree of the parts type. This calculation is locally performed on a common parent part of two parts and parts having a certain information gain is outputted as a cause of failure. How to select these two part attributes is performed in such a way that part attributes located closer to each other on the tree are first evaluated, and first found part attributes are made a candidate of a cause of failure.