Accurate parameter estimation based on latent class model estimated by combining both evaluation and purchase histories

Recently, the recommender system which automatically recommends product items for customers has become more important as an efficient web marketing tool. In many electronic commerce (EC) sites, the data of customers' purchases and evaluation histories are stored in a database. By using them, the system predicts user's preferences, and automatically recommends items that seem to be preferred, but have not been purchased yet. In this study, we focus on the recommender system based on collaborative filtering (CF) with a latent class model. CF recommends items by using purchase or evaluation history data. Considering real purchase activity on EC sites, most of the consumers who bought items on an EC site don't post their evaluation on the site. That means more purchase history data is stored more in the database than evaluation history data. However, most studies of CF used only evaluation data to learn the model. In this case, the purchase data is not used to construct a model even though its data size is much larger than that of evaluation history. It is more desirable to learn a model by using not only evaluation history, but also purchase data to improve the CF accuracy. The purpose of this study is to construct an effective CF model to improve the CF accuracy by formulating the estimation using both evaluation history data and abundant purchase history data which has not been used in previous CF studies. Specifically, we use the aspect model, which is one of latent class models of CF. We propose a way to estimate its parameters using both evaluation history and purchase data. To verify the effectiveness of this study, a simulation experiment was conducted using a bench mark data of recommender system. We show that the prediction accuracy of the recommendation is improved.