With the development of information technology, it has become possible to collect large amounts of customer related data. In particular, retail companies have demands to extract target customers by analyzing the accumulated purchase behavior data of customers. They also need to clarify the impact of items that is related to the good customers. Therefore, a large number of retail companies use their business data for marketing activities. In this study, we focus on a company that make use of a membership-stage system. The membership-stage system sets customer ranks, which we call customer stages, based on the annual spending of each customer. Each customer is given privileges such as discounts depending on their rank. As the customer's rank move up, the customers' purchase behavior varies widely. Therefore, finding the items that lead to for customer's stage growth at each stage becomes an important goal in marketing analysis. For example, there is a framework using a linear classifier to define the importance of items between customers who belong in different stages by Yang et al.. They apply Support Vector Machine(SVM) to the purchase data of customers, classifying whether or not the customers will move up a stage in a year. During the training, moving up a stage is set positive data, and the others are set negative data. Then, they interpret the importance of items using the coefficients of the discrimination function obtained by SVM for each stage. However, the proposal of Yang et al. didn't solve the following two problems. The first problem is that the importance of items among different preferences at each stage could not be interpreted from the results. The second problem is that the features of customers that become a good customer in the first three months of the year will differ from the features of customers in the next nine month, making it difficult to devise effective feature measures. In this study, we propose an analysis method to uncover the impact of latent customer purchase behavior towards stage promotion. We only use three months of customers' purchase data to predict whether the user is capable of becoming a good customer or not using Random Forest. Here, we apply the the PLSA's results of users' purchase behavior for the explanatory variable of Random Forest and interpret the important features in of classes between good customers and the other. Moreover, we apply the proposed method to actual purchase history data to verity the performance of our proposal.