CAECP and CRPD: Classification by Aggregating Essential Contrast Patterns, and Contrast Ranked Path Diagrams

In this paper, we propose a class predictive model for categorical data; this model uses the contrast patterns (CPs) that characteristically appear in a specific class. In a previous study, Morita and Nishiguchi [(2013). Classification model using contrast patterns. In the 15th International Conference on Enterprise Information Systems, Vol. 1, pp. 336–341] proposed a classification model using contrast patterns (CACP). CACP has been applied to practical data and successfully used to construct effective models [Morita, H, Y Shirai and Y Nakamoto (2013). Analysis method for flash marketing incorporating diversification on purchasing. In Annual Conference of Japan Society for Management Information 2013 Autumn, pp. 329–332 (in Japanese); Nishiguchi, M (2014). Classification model using contrast patterns and GRASP. Journal of Information Assurance & Security, 9(5), 235–243]. In their research, we can see that a large number of CPs are required to construct a high-performance model that can be applied to real business data. In terms of knowledge management, the model must provide highly accurate predictions, and offer high readability for their implementations. Such models provide a significant amount of knowledge. To find such knowledge, it is important to select a necessary subset of CPs and to construct a simple model, while maintaining the performance of the model. Further, we propose a visualisation method called Contrast Ranked Path Diagram (CRPD) that allows us to interpret the relationships among the CPs. In computational experiments, we initially apply our method to well-known benchmark problems. From the results, we show that our proposed model outperforms existing methods in terms of the necessary number of CPs and prediction accuracy. Subsequently, we apply our method to actual data from a Japanese retailer, and show the usefulness of the data for business applications.