Abstract
In order to predict performance or reliability of a system including a medical information system, prediction methods using a knowledge base is often used, and various methods have been proposed. However, the accuracy of prediction depends on the characteristics of a system, so finding the best predictor is difficult. In software engineering domain, for example, in error-prone module prediction, this situation is the same, so it is necessary to choose the optimal method for each target system. Zhimin has proposed a selection technique of suitable predictors. In this paper, we modified Zhimin's technique in order to perform mining more appropriately, and we proposed the usage of the features of system characteristics. We performed three experiments based on original data and PROMISE data, and showed the effectiveness our proposal. This technique can be applied to various optimal predictor selection problems.
