Fast Probabilistic Matrix Factorization for recommender system
Wei Feng Yang, Min Tian Wang, Zhou Chen · 2014
In the past decades, with the rapid growth of online user data, it becomes challenging to develop preference learning algorithms that are sufficiently flexible in modeling but also affordable in computation. The enormous datasets and the situation that users who may have few ratings make it extremely hard for many existing approaches to handle. Collaborative filtering[1] is the most successful and popular technology in recommender system. The core of collaborative filtering is collaborative filtering algorithm, while the Probabilistic Matrix Factorization is one of the most useful algorithm. The Probabilistic Matrix Factorization (PMF)[2] model performs well on the large, sparse, and very imbalanced Netflix dataset. However, common methodologies based on error metrics, such as RMSE(Root-Mean-Square Error), are not a natural fit for evaluating the whole recommendation task. In this paper, based on Netflix dataset, we introduce a new Probabilistic Matrix Factorization algorithm called fast PMF, which can get a much better speed results and lower RMSE by updating n% ratings in the top of every movies. By comparing the original PMF, we can get a better understanding of fast PMF. There is a sorting algorithm in fast PMF, we expect a better result by choosing a relatively better sorting algorithm in the future.