27.03.2014 Views

SEKE 2012 Proceedings - Knowledge Systems Institute

SEKE 2012 Proceedings - Knowledge Systems Institute

SEKE 2012 Proceedings - Knowledge Systems Institute

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

An Empirical Study on Recommendation Methods for Vertical B2C E-commerce<br />

Chengfeng Hui, Jia Liu ∗ , Zhenyu Chen, Xingzhong Du, Weiyun Ma<br />

State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China<br />

Software Insititute, Nanjing University, Nanjing, China<br />

∗ liujia@software.nju.edu.cn<br />

Abstract<br />

Recommender systems have been already performing<br />

well in large comprehensive E-commerce sites. Another<br />

trend emerging in E-commerce area is vertical B2C sites.<br />

The vertical B2C sites sell only one or a few of categories<br />

of goods to end users, and most of the users are new users.<br />

There may be not sufficient history data to generate highquality<br />

recommendation because the traffic of these sites are<br />

low. To analyze the feasibility and usability of popular recommendation<br />

methods (e.g. collaborative filtering, contentbased,<br />

etc.) in vertical B2C sites, we have been working in<br />

collaboration with an E-commerce site to gather real data<br />

from an actually running vertical B2C site. In this paper,<br />

we evaluate the performance of different recommendation<br />

methods over half a year period from December 2010 to<br />

June 2011. We analyze both the performance and cost of<br />

these recommendation methods, and experimental results<br />

show that we should apply suitable methods based on the<br />

available data.<br />

Keywords: Vertical E-commerce, Collaborative Filtering,<br />

Content-based Recommendation, Recommender<br />

System<br />

1. Introduction<br />

E-commerce [6] has been widely used to perform business<br />

transactions. One fast growing subcategory of E-<br />

commerce is vertical B2C, and it is becoming more and<br />

more attractive to end users. Unlike mass online merchants<br />

(e.g. Amazon, Walmart), vertical B2C sites focus<br />

on one single or a small number of categories of products<br />

(e.g. Newegg sells computers/electronics, Netflix sells<br />

Books/Music/Movies), they can provide more various and<br />

comprehensive choices in a particular field.<br />

Vertical B2C E-commerce sites are becoming very popular,<br />

but there is little, if any publication directly evaluating<br />

the feasibility of popular recommendation algorithms on<br />

vertical B2C sites, so we want to analyze the performance<br />

of different recommendation methods in such systems. Fortunately<br />

we have the opportunity to collaborate with an E-<br />

commerce site to gather real data from currently running<br />

site and conduct our case study.Vertical B2C E-commerce<br />

sites do have some features that will limit the application<br />

of popular recommendation methods. For content-based<br />

method, the most obvious limitation is items must be capable<br />

of being described as features, if there is no or limited<br />

feature information about items, it is hard for content-based<br />

techniques to get a good performance. For collaborative<br />

filtering, one precondition is there are sufficient ratings to<br />

items given by users or something equivalent. But unlike<br />

data sets (e.g. Movielens, Eachmovie, etc.) which we usually<br />

evaluate collaborative filtering on, it is very likely that<br />

there are no rating information on vertical B2C sites. Another<br />

problem is most users of vertical B2C E-commerce<br />

sites are new users, which will lead to the lack of sufficient<br />

history behavior data of one user to generate his or her profile.<br />

2. Recommender System<br />

The origin of recommender systems can be traced back<br />

to approximation theory, information retrieval and forecasting<br />

theories [1]. The appearance of first papers about<br />

collaborative filtering in the mid-1990s [3] make recommender<br />

systems become an popular research direction.<br />

And recent explosively development of Internet further increases<br />

the interest in this domain both in the industry and<br />

academia. The most famous recommender system applied<br />

in industry is probably the book recommender system of<br />

Amazon (www.amazon.com). This system records users’<br />

purchase, explore, comment and rating data to recognize<br />

users’ preference and then recommends products to users.<br />

The biggest news in academia is maybe in 2009 Netflix<br />

(www.netflix.com) which is a large DVD rental service<br />

company announced to award 1 million dollar prize to a<br />

team that could increase the accuracy of rating prediction by<br />

10%.The BellKor’s Pragmatic Chaos team eventually won<br />

the prize with their BigChaos solution [8].<br />

139

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!