Matic Perovšek (2011) Single-class recommender systems. EngD thesis.
The task of recommender systems is to recommend items that fit the user's preferences. Recommender systems are today often used in web applications and shops in order to help the user in selecting and purchasing items from an overwhelming set of choices. The data from where the hidden preference criteria can be learned often only contains single-class values (web links clicks, bookmarks ...) instead of elaborative ranking. Such data is comprised of only positive examples, listing items that the user has liked or has expressed interest for. For other items, the preference is unknown and may be positive or negative. In this work we study the recommender algorithms that can learn from such data. We examined two types of algorithms. First, RISMF and wALS algorithms, are based on matrix factorization which identifies and later recommends items based on relationships between users and items. We also proposed two types of nearest neighbours algorithms: user and item based. We implemented all of the algorithms and performed a comparison on CiteULike.org website’s database. Results show that wALS algorithms give the best results on the selected data.
Actions (login required)