Distributed randomized singular value decomposition using count sketch

Compared with other recommendation algorithms, Matrix decomposition is frequently used in the current recommendation system. It can not only lead to better results, but also can fully take the influence of various factors into account, which explains its good scalability. Matrix decomposition includ-es SVD(Singular Value Decomposition), non-negative matrix decomposition, Latent Factor Model and some other traditional matrix decomposition techniques is designed to approximate a high-dimensional matrix with low-dimensional. As a perfect technique in recommendation system, SVD is traditionally expert at dense matrix decomposition. However, real rating matrix are sparse, and have high time complexity of SVD, if the matrix size increases rapidly, the efficiency must become unacceptable. The combination of random algorithm and matrix decomposition turns traditional matrix decomposition into random matrix decomposition technique under distributed system environment. The random singular value decomposition technique illustrated in the following content can be at the expense of little accuracy under the premise of greatly improving the efficiency of the calculation.