CSMC

CSMC is a Python library for performing column subset selection in matrix completion tasks. It provides an implementation of the CSSMC method, which aims to complete missing entries in a matrix using a subset of columns.

Columns Selected Matrix Completion (CSMC) is a two-stage approach for low-rank matrix recovery. In the first stage, CSMC samples columns of the input matrix and recovers a smaller column submatrix. In the second stage, it solves a least squares problem to reconstruct the whole matrix.

CSMC supports numpy arrays and pytorch tensors.

Installation

You can install CSMC using pip:

pip install -i  csmc

Usage

1. Generate random data

import numpy as np
import random 

n_rows = 50
n_cols = 250
rank = 3

x = np.random.default_rng().normal(size=(n_rows, rank)) 
y = np.random.default_rng().normal(size=(rank, n_cols)) 
M = np.dot(x, y)

M_incomplete = np.copy(M)
num_missing_elements = int(0.7 * M.size)
indices_to_zero = random.sample(range(M.size), k=num_missing_elements)
rows, cols = np.unravel_index(indices_to_zero, M.shape)
M_incomplete[rows, cols] = np.nan

2. Fill with CSNN algorithm

from csmc import CSMC
solver = CSMC(M_incomplete, col_number=100)
M_filled = solver.fit_transform(M_incomplete)

3. Fill with Nuclear Norm Minimization with SDP (NN algorithm)

from csmc import NuclearNormMin
solver = NuclearNormMin(M_incomplete)
M_filled = solver.fit_transform(M_incomplete, np.isnan(M_incomplete))

3. Fill with Frank-Wolfe (Conditional Gradient Method)

from csmc import CGM
solver = CGM(M_incomplete)
M_filled = solver.fit_transform(M_incomplete, np.isnan(M_incomplete))

Algorithms

• NuclearNormMin: Matrix completion by SDP (NN algorithm) Exact Matrix Completion via Convex Optimization
• CSNN: Matrix completion by CSNN
• PGD: Nuclear norm minimization using Proximal Gradient Descent (PGD) Spectral Regularization Algorithms for Learning Large Incomplete Matrices by Mazumder et. al.
• CSPGD: Matrix completion by CSPGD
• CGM: Matrix completion with Frank-Wolfe method
• SGD: Scaled Gradient Descent (SGD) [Accelerating Ill-Conditioned Low-Rank Matrix Estimation via Scaled Gradient Descent(https://jmlr.org/papers/volume22/20-1067/20-1067.pdf) by Tong et. al.
• SVP: Singular Value Projection ( SVP) Guaranteed Rank Minimization via Singular Value Projection by Meka et. al.
• MC2: MC2 MC2: a two-phase algorithm for leveraged matrix completion by Eftekhari et. al.
• CUR+Nuc: CUR+Nuc based on CUR+ for uniformly sampled entries CUR Algorithm for Partially Observed Matrices by Xu et. al.

Examples

• small synthetic matrices
• big synthetic matrices
• small images inpainting
• big images inpainting
• cgm

Configuration

To adjust the number of threads used for intraop parallelism on CPU, modify variable:

NUM_THREADS = 8

in settings.py

Citation

Krajewska, A., Niewiadomska-Szynkiewicz E. (2024). Randomized Approach to Matrix Completion: Applications in Recommendation Systems and Image Inpainting.

Krajewska, A. , Niewiadomska-Szynkiewicz E., "Efficient Data Completion and Augmentation." 2024 IEEE 11th International Conference on Data Science and Advanced Analytics (DSAA). IEEE, 2024.