In the retail service, knowledge management with point of sales (POS) data mining is integral to maintaining and improving productivity. The present paper describes a method of computational customer behavior modeling based on real datasets, and we demonstrate some knowledge extractions from the model. The model is constructed by Bayesian network based on a large-scale POS dataset that incorporates customer identification information and questionnaire responses. In addition, we employ an automatic categorization using probabilistic latent semantic indexing (PLSI), because an appropriate categorization of customers and items is needed for construction of a useful model in real services. We identify a number of categories with regard to customer behavior, and demonstrate the efficacy of our knowledge extraction approach for effective service provision and knowledge management.