Efficacy and safety in randomized controlled clinical trials are often evaluated between active (treatment) and control groups. A limitation is that there may be disparity in (average) treatment effect between subjects. Statistical approaches to determine subgroups with higher treatment effect so to distinguish subgroups have therefore been developed.
　Threshold-based methods, such as CART (Classification and Regression Trees; Breiman et al., 1984), have been applied as methods of subgroup identification. Dusseldorp & Mechelen (2013) proposed the `QUINT' method, which uses Cohen's effect size as evaluation criteria. Meanwhile, tree-structured methods are a statistical approach, in which treatment effects are classified into similar subgroups, and are therefore not necessarily a suitable method for identifying responders.
　The present study instead focuses on the PRIM method (Patient Rule Induction Method; Friedman & Fisher, 1999). Previously reported PRIM-based methods (SPRIM; Kehl & Ulm, 2006) assumed proportional hazards between active (treatment) and control groups.
　We consider the application of a different PRIM method for survival data without such assumption of proportional hazards. It applies restricted mean survival time (RMST) to evaluate the measurement of treatment effect and is used to establish the subgroups to minimize or maximize any differences in treatment effect. The effectiveness of this method is illustrated through a practical example; a small-scale simulation suggests that it can be used to identify more appropriate subgroups than the SPRIM and SIDES (Lipkovich et al., 2011) methods.

View full abstract

　In this study, we analyzed two years of customer information and point-of-sale data for several stores in a hair salon chain; we proposed how to improve their sales. First, to classify the characteristics of customer service use, we employ RFM (Recency, Frequency, Monetary) analysis. Then, we propose the metric (hourly) by studying the customer's accounting amount per hour because the time varies depending on the service. We suggest the classification of RFM (hourly) as “RFM+H analysis,” and analyze the purchase tendency of customers for each store. Furthermore, using RFM+H analysis, we redefine the customer classification. With this classification, we analyze the relationship between the customers' addresses and the stores they patronize, and the change the customer classification of the customer by two terms. Finally, we create a Shiny application that interactively visualizes our analysis, using RStudio, and considered it.

View full abstract

　In this paper, we visualized and predicted customer information and accounting information included in ID-POS data of the hair salon chain. Our purpose was to increase the number of repeat customers.
　We described our software developed for the purpose and its results. We grasped trends of customer information and accounting information by visualizing the ID-POS data. As the results, we grouped the customers. The group to which customers belong is predicted from their accounting information. Specific products were recommended to one group of customers, while the next visiting dates were predicted for another group of customers. We predicted customer groups using deep learning, and searched recommended products using collaborative filtering, respectively.

View full abstract

　In managing relationships with customers, factors contributing to the frequency of visits to stores and customers' dropouts are important. In this research, we introduce the Markov renewal process, in which the occurrence of an event will be stopped with a certain probability. The conventional Pareto/negative binomial distribution (NBD) model cannot express the distribution of customers' visiting intervals when their densities are concentrated around the mode. We propose a more flexible model that can estimate the coefficients and confidence intervals of variables that affect the probability of customers' dropouts and their visiting frequency. The biases and standard errors of the estimators are evaluated using a simulation study. The usefulness of the proposed model is indicated based on the analysis of customer data from a hair salon.

View full abstract