Mobile phone market is in mature in Japan and mobile phone carriers struggle to prevent their users from outflowing to other mobile phone companies. Mobile phone companies develop various marketing strategies to hold good customers, for example, discounting communication charge drastically or delivering discount coupon ticket for long-term users. To identify good customers, they use cancellation prediction model based on demographic data like age and sex and transaction data like mobile airtime. By the way, mobile phone companies can use various types of data within the related law, but it is also fact that they cannot use those data thoughtfully yet because those data is very big and computer resource is limited. In this study, we pay attention to search word put into search engine and propose how to make cancellation prediction model of mobile phone with use of search word.

This paper evaluates the performance of Flexible Mixture Model which is one of the probabilistic latent variable models for recommender system. Recently, previous researches showed that the performances of memory-based collaborative filtering methods can be evaluated by Agent-based simulation. Though the effectiveness of methods for recommender system was sometimes clarified by using benchmark data sets, the technique of Agent-based simulation gives us the way to investigate its’ properties from several different viewpoints. It was meaningful to investigate the performance and behavior of memory-based collaborative filtering methods. However, several model-based collaborative filtering methods have been proposed and their effectiveness has been pointed out. It is meaningful to inves- tigate the characteristics and performance of modelbased collaborative filtering. This paper focuses on Flexible Mixture Model and investigate its’ characteristics by Agent-based simulation techniques.