An initialization method to improve the training time of matrix factorization algorithm for fast recommendation
- PDF / 1,825,237 Bytes
- 13 Pages / 595.276 x 790.866 pts Page_size
- 18 Downloads / 176 Views
(0123456789().,-volV)(0123456789(). ,- volV)
METHODOLOGIES AND APPLICATION
An initialization method to improve the training time of matrix factorization algorithm for fast recommendation Mojtaba Mohammadian1 • Yahya Forghani1
•
Masood Niazi Torshiz1
Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract Recommendation systems are successful personalizing tools and information filtering in web. One of the most important recommendation methods is matrix factorization method. In matrix factorization method, the latent features of users and items are determined in such a way that the inner product of the latent features of a user with the latent features of an item is equal to that user’s rating on that item. This model is solved using alternate optimization algorithm. The solution and the prediction error of this algorithm depend on the initial values of the latent features of users which are usually set to small random values. The purpose of this paper is to propose a fast alternate optimization algorithm for matrix factorization which converges to a good solution. To do so, firstly, we show experimentally that if the latent feature vector of each user is initialized by a vector of which elements are equal, we can also obtain a proper solution using the alternate optimization algorithm. Then, we prove that if our proposed initialization method is used, the alternate optimization algorithm for matrix factorization can be simplified using Sherman–Morrison formula. Experimental results on 5 real datasets show that the runtime of our proposed algorithm is 2–45 times less than the traditional method. Keywords Matrix factorization Recommendation system Initialization method Sherman–Morrison formula
1 Introduction Recommendation systems are useful tools for information filtering and are widely used in commercial applications such as playlist generators for music and video services like YouTube and Netflix, content recommenders for social media services such as Facebook and Twitter, or product recommenders for platforms such as Amazon. Matrix factorization is one of the most famous collaborative filtering recommendation methods. Matrix factorization method (MF) (Yuan et al. 2018) learns latent features of users and items in such a way that the inner product of the latent features of a user with the latent feature of an item is equal Communicated by V. Loia. & Yahya Forghani [email protected] Mojtaba Mohammadian [email protected] Masood Niazi Torshiz [email protected] 1
to that user’s rating on that item. Then, the inner product of the latent features of a user with the latent features of an item is used to predict that user’s rating on that item. MF is solved using alternate optimization algorithm of which convergence has been guaranteed (Bertsekas 2016). In this approach, first, latent feature vectors of users are initialized and then, in each iteration, first latent feature vectors of items are updated, and then latent feature vectors of users are updated until converge
Data Loading...
