Recommending content using side information
- PDF / 3,137,971 Bytes
- 22 Pages / 595.224 x 790.955 pts Page_size
- 42 Downloads / 238 Views
Recommending content using side information Rabeh Ravanifard1 · Wray Buntine2 · Abdolreza Mirzaei1 Accepted: 12 September 2020 © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Collaborative Filtering methods predict user interests and make recommendations just by using the rating matrix. However, in practice there is extensive side information about users and items, such as the age of the user, the actors in a movie, or the abstract of a journal article. In this paper, a novel model called Collaborative Poisson Factorization with Sideinformation (CPFS) is proposed which extends CTPF by incorporating richer kinds of side information conditionally as a prior to the model. CPFS is a monolithic hybridization model that combines features from different data sources into a single recommendation algorithm. We develop a Gibbs sampler and also a Variational method with closed-form updates for the inference of CPFS and demonstrate its applicability on a range of datasets including movies, books, academic papers, and travel. The extension improves prediction quality, especially in the cold start scenario. The connections between side information and topics are also intuitive. Keywords Poisson matrix factorization · recommender systems · side information
1 Introduction Recommender systems [21, 28, 33] try to predict users’ interests and recommend a list of items to the users. One of the most significant techniques for recommender systems is Collaborative Filtering (CF) that make prediction based on the similarity between users or items and without any additional information. CF methods use neighbourhood methods or latent factor models to measure the similarity [21]. Matrix Factorization (MF) models are among the most used types of latent factor models that perform well for recommender systems [21, 38]. MF methods model the user preference matrix as a product of lower-rank user and item matrices, often referred to as latent factors.
Abdolreza Mirzaei
[email protected] Rabeh Ravanifard [email protected] Wray Buntine [email protected] 1
Department of Electrical and Computer Engineering, Isfahan University of Technology, Isfahan, 84156-83111, Iran
2
Department of Information Technology, Monash University, Melbourne, Australia
The recommendation accuracy of CF and consequently MF methods can be strongly affected by the sparsity of available ratings. They need a sufficient number of ratings, otherwise the recommendation performance is reduced significantly [6]. Furthermore, these methods suffer from the cold-start problem that occurs when a new user or a new item is entered into a recommender system: the system does not have any historical data about them, so the prediction will not be accurate. Despite these limitations, one strength of MF methods is their flexibility to integrate additional information, sometimes referred to as side information. Therefore, over the past few years, various approaches have been proposed to relieve the cold-start problem by incorporating additio
Data Loading...
