Abstract
Information filtering is an area getting more important as we have long been flooded with too much information, where product brokering in e-commerce is a typical example. Systems which can provide personalized product recommendations to their users (often called recommender systems) have gained a lot of interest in recent years. Collaborative filtering is one of the commonly used approaches which normally requires a definition of user similarity measure. In the literature, researchers have proposed different choices for the similarity measure using different approaches, and yet there is no guarantee for optimality. In this paper, we propose the use of machine learning techniques to learn the optimal user similarity measure as well as user rating styles for enhancing recommendation acurracy. Based on a criterion function measuring the overall prediction error, several ratings transformation functions for modeling rating styles together with their learning algorithms are derived. With the help of the formulation and the optimization framework, subjective components in user ratings are removed so that the transformed ratings can then be compared. We have evaluated our proposed methods using the EachMovie dataset and succeeded in obtaining significant improvement in recommendation accuracy when compared with the standard correlation-based algorithm.
Original language | English |
---|---|
Pages (from-to) | 395-410 |
Number of pages | 16 |
Journal | Information Retrieval |
Volume | 7 |
Issue number | 3-4 |
DOIs | |
Publication status | Published - 2004 |
Scopus Subject Areas
- Information Systems
- Library and Information Sciences
User-Defined Keywords
- Collaborative filtering
- Machine learning
- Rating style
- Recommender systems
- User similarity