USER INTENT PREDICTION FROM ACCESS LOG IN ONLINE SHOP

A lot of recommendation systems on online shops use user's order histories in order to determine recommendation items. In general recommendation systems items are selected based on neighbor users defined according to similarity among users using the order histories. However, the method cannot be applied to new users who have never purchased anything in an online shop and do not define the neighbor users based on their order histories because of undefined similarity. The problem is called a cold start problem. In order to overcome the problem we proposed a method which uses user's access logs to make user profile instead of his/her order histories. Although the access log is less reflective of the user’s preference than the order history, we can estimate their intent by careful access log analysis because the access log consists of user’s review processes for their purchase. Therefore, to clarify users' intent we use only web pages related to decision of order products strongly. And in order to find these web pages we analyze access logs of users who ordered the same product or the same category. Then we use these pages for making a user profile. In experiments we estimate neighbor users of new users using their user profiles constructed with access logs. And we predict a category of a product which the new users will purchase to examine the efficiency of our proposed method. From experiments we found that there were some categories in which the proposed method can correctly predict new user’s target, we confirmed the effectiveness of our proposed method as a solution for cold start problem. However, we found that we improved the proposed method to predict a product itself.