In online shops, many recommendatio

In online shops, many recommendation systems are used to improve a conversion rate
(Kamishima et al 2006). The conversion rate is a purchasing rate for visitors. If we can
recommend products that users are requiring or searching, we achieve a high conversion rate.
In general recommendation products are selected based on neighbor users defined according to
similarity among order histories or users’ demographic information registered previously
(Linden et al. 2003, Auer et al. 2002 and Gittins et al. 2011). The neighbor users are users who
are considered to have a similar preference to a target user. Hence, it is important to find
correct neighborhood in the systems. However, there is a problem that we cannot define
neighborhood for new users who use the site for the first time or have never purchased any
products since we cannot define similarities between the new user and another user. This
problem is called a cold start problem (Sahebi 2011, Zhang et al 2010).
In this paper, we overcome the cold start problem using access logs in an online shop
instead of order histories since the access logs are obtained more easily. Our aim is to define
appropriate neighbor users from access logs and predict new users’ intents. However, access
logs are not reflective of the user’s preference directly, compared with the order history. For
example, a user just browses various products according to temporary interests in some cases
and compares a few products to decide what to buy according to some concrete interests. So
the access logs include some access logs not related to user’s purchase as noise. Then, we pay
attention to access logs of user’s review processes on their purchasing, since the processes
sure to have specific intents. And in order to capture user’s various intents using web pages we
pay attention to common pages in access logs of users who ordered the same product or the
same category. This is because that we consider that web pages visited by many customers on
their purchasing processes have an effect on the purchasing of the category-specific.
Concretely, we construct a network from the access logs. Evaluating the network, we obtain
important web pages which are highly relevant to customers’ purchase. In this paper, we use
PageRank algorithm (Page et al 1998) to select these web pages and call them
“Characteristics Pages; CPs” (Koketsu et al 2012). Using CPs in each category, we make a
feature vector which elements denote whether the user visited the CPs or not as the user’s
profile. In experiments we estimate neighbor users of new users by calculating similarities
among users’ profiles. And we predict the category of product which the new users will
purchase to examine the efficiency of our proposed method. From experimental results in
category prediction we found that there were some categories in which the proposed method
can correctly predict the product category that new users actual purchased, we confirmed the
positive effectiveness of the proposed method as one solution for the cold start problem.