IncSVD: A python package for dynamic Truncated SVD

IncSVD is a python package capable of dynamically maintaining the incremental Truncated Singular Value Decomposition (Truncated SVD) of evolving matrices.

This is the code implementation of the paper "Fast Updating Truncated SVD for Representation Learning with Sparse Matrices" (ICLR'24).

Install and Setup.

It is recommended to use python 3.8 or higher and install IncSVD with the following command.

pip install IncSVD

User Guide

Initialize the IncSVD using an $m \times n$ matrix data_init given the rank (dimension) k for which the truncated SVD needs to be computed. Here is an example to initialize a $1000 \times 1000$ matrix with truncated dimension $k = 64$.

class EvolvingMatrix(data, k, sparse = True, method = "RPI")

• data: data matrix Data matrix used to initialize truncated SVD. The format of this initial matrix can be np.ndarray or scipy.sparse.
• k: int Truncated dimension in the maintained SVD. It should be a positive integer not greater than the length of the sides of the original matrix.
• sparse: bool, default=True Set true implies the update process will employ the proposed acceleration methods [1]. Set false implies using the original methods (ZhaSimon's [2], Vecharynski's [3], Yamazaki's [4]).
• method: {'ZhaSimon', 'GKL', 'RPI'}, default='RPI' Method for SVD update. It can be chosen from ['ZhaSimon', 'GKL', 'RPI'].

import scipy.sparse
import IncSVD.EvolvingMatrix as EM
m, n, k = 1000, 1000, 64

# data_init = scipy.sparse.rand(m, n, density=0.005).tocsr()
M = EM.EvolvingMatrix(data_init, k, sparse=True, method="ZhaSimon")

Row / Column Update

Expand the original matrix in row or column dimensions and update the corresponding SVD results.

add_row(update_matrix, l=None, t=None)

• update_matrix: data matrix Data matrix to be appended. This matrix should have the same number of columns as the original matrix. If type sparse is True, this matrix should be in scipy.sparse._csr.csr_matrix.
• l: {int, None} Approximate dimension for the update matrix. When using GKL or RPI, two methods that require an approximation space, it is necessary to set l. The update matrix of rank s will be approximated with a smaller space of rank l for further speedup. Set to None when GKL and RPI is not applied.
• t: {int, None} The number of iterations to perform in the random power iteration. It needs to be set when using the RPI. This approximation is more accurate when t is larger, while the computational overhead is larger. Set to None when RPI is not applied.

add_column(update_matrix, l=None, t=None)

• update_matrix: data matrix Data matrix to be appended. This matrix should have the same number of rows as the original matrix. If type sparse is True, this matrix should be in scipy.sparse._csc.csc_matrix.
• l: {int, None} Approximate dimension for the update matrix. Set to None when GKL and RPI is not applied.
• t: {int, None} The number of iterations to perform in the random power iteration. Set to None when RPI is not applied.

''' Add rows '''
# data_append_row = scipy.sparse.rand(m, 10, density=0.005).tocsr() * 10
M.add_row(data_append_row)

''' Add columns '''
# data_append_col = scipy.sparse.rand(n, 10, density=0.005).tocsc() * 10
M.add_column(data_append_col)

Low-rank Update.

Perform low-rank update $\overline{\mathbf{A}} = \mathbf{A} + \mathbf{D}_m \cdot \mathbf{E}_m^{\top}$, where $\mathbf{A}$ is the original matrix, $\mathbf{D}_m \in \mathbb{R}^{m \times s}$ and $\mathbf{E}_m \in \mathbb{R}^{n \times s}$.

update_weight(self, update_matrix_B, update_matrix_C, l=None, t=None):

• update_matrix_B: data matrix Data matrix to be appended. This matrix should have the same number of columns as the original matrix. If type sparse is True, this matrix should be in scipy.sparse._csc.csc_matrix.
• update_matrix_C: data matrix Data matrix to be appended. This matrix should have the same number of columns as the original matrix. If type sparse is True, this matrix should be in scipy.sparse._csc.csc_matrix.
• l: {int, None} Approximate dimension for the update matrix. Set to None when GKL and RPI is not applied.
• t: {int, None} The number of iterations to perform in the random power iteration. Set to None when RPI is not applied.

''' Low-rank update '''
# data_dm = scipy.sparse.rand(m, 10, density=0.005).tocsc() * 10
# data_em = scipy.sparse.rand(n, 10, density=0.005).tocsc() * 10
M.update_weight(data_dm, data_em)

Get the result.

Queries the current SVD result.

''' Get the result. '''
Uk, Sigmak, Vk = M.Uk, M.Sigmak, M.Vk

''' Get a row of singular vectors. '''
# i = 0
Uk0, Vk0 = M.Uki(i), M.Vki(i)

''' A low-rank approximation of data matrix can be obtained from. '''
Ak = Uk @ np.diag(Sigmak) @ Vk.T

Contact Us

If you have any questions about this code repository, feel free to raise an issue or email denghaoran@zju.edu.cn.

Citations.

If you find our work useful, please consider citing the following paper:

@inproceedings{deng2024incsvd,
title={Fast Updating Truncated {SVD} for Representation Learning with Sparse Matrices},
author={Haoran Deng and Yang Yang and Jiahe Li and Cheng Chen and Weihao Jiang and Shiliang Pu},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=CX2RgsS29V}
}