BoostMF: Boosted Matrix Factorisation for Collaborative Ranking
Personalised recommender systems are widely used information filtering for information retrieval, where matrix factorisation (MF) has become popular as a model-based approach to personalised recommendation. Classical MF methods, which directly approximate
- PDF / 378,765 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 82 Downloads / 163 Views
Abstract. Personalised recommender systems are widely used information filtering for information retrieval, where matrix factorisation (MF) has become popular as a model-based approach to personalised recommendation. Classical MF methods, which directly approximate low rank factor matrices by minimising some rating prediction criteria, do not achieve a satisfiable performance for the task of top-N recommendation. In this paper, we propose a novel MF method, namely BoostMF, that formulates factorisation as a learning problem and integrates boosting into factorisation. Rather than using boosting as a wrapper, BoostMF directly learns latent factors that are optimised toward the top-N recommendation. The proposed method is evaluated against a set of stateof-the-art methods on three popular public benchmark datasets. The experimental results demonstrate that the proposed method achieves significant improvement over these baseline methods for the task of top-N recommendation. Keywords: Recommender system · Collaborative filtering · Matrix factorisation · Learning to rank · Boosting
1
Introduction
Recommender systems (RS) have gained much attention in information retrieval (IR) to guide users when searching information from the information pool. Collaborative filtering (CF) is widely used to build personalised recommender systems such as book recommendation in Amazon [2], movie recommendation in Netflix [2] and friend recommendation in Facebook [2]. It aims to predict the preference of a user on its unseen items by learning the preference from the historic feedback of this user and other like-minded users to provide the user with a list of recommended items or prediction score of items. The personalised prediction problem [1–3,15] in presenting recommendation list can be regarded as estimating the preference function in CF. Usually, this problem can be solved by either i) generating the recommendation list by sorting the predicted ratings in descending order, known as rating-oriented CF or ii) learning the ranking function directly, known as ranking-oriented CF. When the recommendation list itself becomes large, it will be obsolete since people prefer only top listed items c Springer International Publishing Switzerland 2015 A. Appice et al. (Eds.): ECML PKDD 2015, Part II, LNAI 9285, pp. 3–18, 2015. DOI: 10.1007/978-3-319-23525-7 1
4
N. Chowdhury et al.
[2,15]. So recommender systems should not only be optimised to reflect user tastes and preferences but also rank top items correctly. Matrix factorisation is a popular model-based CF method, which demonstrates great success in Netflix prize competition [7]. In MF, given N users and M items, the user-item preference matrix R ∈ N ×M can be approximated by two low rank matrices P ∈ N ×K and Q ∈ M ×K as R ≈ P · Q by minimising the sum of squared errors, where K min(N, M ) is the dimensionality of latent factors representing user preferences and item characteristics. The major purpose of MF is to obtain some forms of lower-rank approximation to original matrix for unde
Data Loading...
