PROJECT TITLE :

Distributed Bayesian Matrix Decomposition for Big Data Mining and Clustering

ABSTRACT:

Matrix decomposition is one of the essential tools that must be utilized in order to extract useful information from the massive amounts of data that are produced by contemporary applications. To process extremely large datasets utilizing such a method on a single machine is, despite this, either inefficient or impossible to achieve. Additionally, Big Data are frequently collected and stored in a dispersed fashion across multiple computers. Therefore, such data typically contain a significant amount of noise that is of a heterogeneous nature. For the analysis of big data, developing a distributed matrix decomposition is something that is both necessary and helpful. A method like this should have good scalability, be able to model heterogeneous noise, and provide an answer to the problem of Communication in a distributed system. In order to achieve this goal, we propose a distributed Bayesian matrix decomposition model (DBMD) for the mining and clustering of large amounts of data. In particular, we use the accelerated gradient descent, the alternating direction method of multipliers (ADMM), and statistical inference to implement the Distributed Computing. These three strategies are: 1) the alternating direction method of multipliers; 2) the accelerated gradient descent; and 3) the statistical inference. The theoretical convergence behaviors of these algorithms are investigated by us. We propose an optimal plug-in weighted average that can reduce the variance of the estimation in order to deal with the fact that the noise comes from a variety of sources. Real-world experiments demonstrate that our algorithms achieve superior or competitive performance when compared to two typical distributed methods, including scalable k-means++ and Scalable-NMF. Synthetic experiments validate our theoretical results. Real-world experiments show that our algorithms scale up well to big data.