Neighborhood User Estimation fromWeb Access Log in EC Service

An aim of this paper is to find neighborhood users using customers´ access logs in an Electric Commerce site. In general recommendation services neighborhood users usually are defined according to their order histories or their demographic information. Since a neighborhood user estimation algorithm does not define the neighborhood of users that have never bought any products in an EC site, a cold start problem happens in collaborative filtering. To overcome the cold start problem we make user profile from his/her access logs instead of order histories. We have to select access logs that show user´s intent clearly since access logs include user´s activities that are not related to product purchase. Hence, we focus on Web pages that many users who purchase the same products visited. We assume the Web pages affect user´s purchase and use them as characteristic Web pages to predict users´ intent. Since we think such the characteristic Web pages are different in product categories we define a set of Web pages in each category. After finding the characteristic Web pages in all categories, we make a feature vector which elements denote the characteristic Web pages and value denotes whether the user visited the Web pages or not. We calculate the similarities among product categories using sets of users buying a product included in a category to check a user profile discriminates categories. Carrying out evaluation experiments, we confirmed that there is high similarity among users who purchase products in the same category and the similarity is a discriminative criterion. And we estimate neighborhood of new users, which have never bought any products, and confirm that the neighborhood includes many users that have bought products in the same category.