A Recommender System with Interest-Drifting
Collaborative filtering and content-based recommendation methods are two major approaches used in recommender systems. These two methods have some drawbacks in dealing with situations such as sparse data and cold start problems. Recently, combined methods
- PDF / 362,989 Bytes
- 10 Pages / 430 x 660 pts Page_size
- 69 Downloads / 237 Views
2
School of Information Technology and Electrical Engineering, University of Queensland, QLD 4072, Australia Polish-Japanese Institute of Information Technology,Faculty of IT, Ul. Koszykowa 86, 02-008 Warsaw, Poland {shanle, xueli,ding,maria}@itee.uq.edu.au
Abstract. Collaborative filtering and content-based recommendation methods are two major approaches used in recommender systems. These two methods have some drawbacks in dealing with situations such as sparse data and cold start problems. Recently, combined methods were proposed to overcome these problems. However, a highly effective recommender system may still face a new challenge on interest drift. In this case, customer interests may change over time. For example, more recent users’ ratings on items may reflect more on users’ current interests than those of long time ago. Unfortunately, current available combination approaches do not consider this important factor and training data sets are regarded as static and time-insensitive. In this paper, we present a novel hybrid recommender system to overcome the interest drift problem by embedding the time-sensitive functions into the recommendation process. The users’ interests changing behaviours are considered with time function. Our experiments demonstrate a better performance than that of the collaborative filtering approaches considering interests drift and those of the combined approaches without considering interests drift.
1
Introduction
Recommender systems were introduced to help users to overcome information overload problem and have been widely used in e-commerce websites. Most prevalent recommender systems apply either content-based filtering or collaborative filtering approaches to predictions. Content-based filtering approach analyzes the similarity between items based on their contents and recommends the similar items based on users’ previous preferences. On the other hand, collaborative filtering approach computes the similarity between users according to their historical ratings to items. Collaborative filtering systems recommend user with the items that were rated highly in the past based on a ranking approach of finding nearest neighbours [1]. Also, both content-based filtering and collaborative filtering approaches have shortcomings and perform badly in some situations. For example, for contentbased approach, it needs to analyze profiles of items and then compute the similarity between items’ profiles that were purchased or preferred by users as well as the profiles of items that are not rated or purchased. But profiles are hard B. Benatallah et al. (Eds.): WISE 2007, LNCS 4831, pp. 633–642, 2007. c Springer-Verlag Berlin Heidelberg 2007
634
S. Ma et al.
to extract from some data types such as audio/video products. Another problem inherent in content-based filtering approach is about the limitation of the scope of possible interests. The interesting items that system can recommend are only those, for which users have preferred or purchased. This is regarded as a “coldstart” problem [2]. Furthermore,
Data Loading...
