Web Mining

Investigators

Obradovic Zoran
Vucetic Slobodan

Problem

In today's society there is an increasing need for automated systems providing personalized recommendations to a user faced with a large number of choices. The products customization trend coupled with E-commerce where customers were not provided with an option to examine the products “off-shelf” in a traditional sense, make the problem of providing accurate personalized recommendations very important. Automated methods are needed to provide a large number of users with the ability to efficiently locate and retrieve information according to their preferences. The task of collaborative filtering is to predict preferences of an active user on unseen items given preferences of other users, typically expressed as numerical ratings.

Results

We proposed [vucetic00d] a novel regression-based approach that learns a number of experts describing relationships in ratings between pairs of items. Based on ratings provided by an active user for some of the items, the experts are combined efficiently to predict user’s preferences on the remaining items. Extensive experiments on Eachmovie and Jester benchmark collaborative filtering data showed that the proposed regression-based approach achieves improved accuracy and is orders of magnitude faster than the popular neighbor-based alternative.