Research On Matrix Factorization-based Recommendation Algorithm

In the age of big data, the amount of information on the Internet has increased so much that makes it difficult for users to obtain the interesting information. Recommendation system is an important tool to provide users with the items they really need, and it is also useful to overcome information overload problem. The matrix factorization based recommendation algorithm is the state-of-art recommendation method. In matrix factorization based algorithm, user behavior matrix is factorized as the product of two matrices in the latent factor space. Even though this kind of algorithm obtains the promising performance in recommendation area, there still are some problems, for example it can’t deal with the one-class data and sparse data very well. Meanwhile, they are faced by the problem of big data recommendation because of high computational complexity.In this paper, we focus on the matrix factorization based recommendation algorithms and propose some improved algorithms to solve the problems mentioned above. The main contributions are summarized as follows.Firstly, we proposed an improved matrix factorization based recommendation algorithm for one-class problem. Usually, recommendation can be taken as a problem of one-class data classification. In this case, however, the sparsity of data and the lack of negative samples always affect the final recommendation result. In this paper, we aimed to choose positive and negative samples by considering both item similarity and user activity, so that the influence of the sparsity and the lack of negative samples can be effectively reduced. A series of experiments were conducted and the results have shown that the new method can significantly improve the recommend accuracy.Secondly, a sparse probabilistic matrix factorization based recommendation algorithm is presented. Probabilistic matrix factorization can provide good interpretation for latent factors from a probabilistic perspective. However, the existing probabilistic matrix factorization models including Bayesian probabilistic matrix factorization usually assume that user/item feature factors follow Gaussian distribution, which is not accurate when data are very sparse. To solve this problem, we put forward a new algorithm by using Laplace distribution to replace Gaussian distribution, which can characterize the spare data effectively. A series of experiments on Netflix and MovieLens datasets have been conducted to demonstrate that the sparse probabilistic matrix factorization model outperforms the existing methods, especially on much sparse dataset (e.g., Netflix). Meanwhile, Laplace distribution is long tail distribution, which makes the proposed method more suitable to handle the long tail items.Finally, we designed a parallel algorithm to optimize the L1-regularized matrix factorization model. In this chapter, we theoretically prove that L1-regularized matrix factorization model has the same probabilistic interpretation with sparse probabilistic matrix factorization. Considering the complexity of the inferring process to solve the probabilistic matrix factorization, we focus on solving the L1-regularized matrix factorization model in the framework of Map-Reduce.