One of the most important problems in data mining is how to manage a large amount of data and to extract efficient knowledge from large databases. Although many machine learning methods and statistical methods have been proposed to solve this problem, they are not powerful when we have more than 1000 samples, since the computational complexity of these algorithms is larger than or approximately equal to n2. In this paper, we introduce the idea of log-likelihood ratio to measure the similarity between generated training samples and original training samples before rule induction methods are applied to this selected samples. This method was evaluated to three medical domains. The results show that the proposed method selects training samples which reflect the statistical characteristics of the original training samples although the performance with small samples is not so good.