Applications of Online Nonnegative Matrix Factorization to Image and Time-Series Data

Hanbaek Lyu,Georg Menz,Deanna Needell,Christopher Strohmeier

DOI: https://doi.org/10.48550/arXiv.2011.05384

2020-11-11

Abstract:Online nonnegative matrix factorization (ONMF) is a matrix factorization technique in the online setting where data are acquired in a streaming fashion and the matrix factors are updated each time. This enables factor analysis to be performed concurrently with the arrival of new data samples. In this article, we demonstrate how one can use online nonnegative matrix factorization algorithms to learn joint dictionary atoms from an ensemble of correlated data sets. We propose a temporal dictionary learning scheme for time-series data sets, based on ONMF algorithms. We demonstrate our dictionary learning technique in the application contexts of historical temperature data, video frames, and color images.

Machine Learning

What problem does this paper attempt to address?

The problems that this paper attempts to solve are: in the data stream environment, how to effectively use the Online Nonnegative Matrix Factorization (ONMF) algorithm to learn joint dictionary atoms from a large number of related data sets and apply them to the processing of image and time - series data. Specifically, the paper explores the following aspects:

1. **Applications of Online Nonnegative Matrix Factorization**:
- How to update the dictionary and encoding matrix in real - time in the data stream environment to adapt to newly arrived data samples.
- How to learn common feature representations (i.e., joint dictionary atoms) from multiple related data sets through the ONMF algorithm.

2. **Time - series data processing**:
- Use ONMF for online dictionary learning and online reconstruction, especially to reveal the temperature dependence between cities in historical temperature data and use it to fill in missing data.
- Detect significant changes in time - series by factorizing the time dimension.

3. **Image processing applications**:
- Use ONMF for image compression, reducing storage requirements by learning low - rank representations of small image patches.
- Combine Convolutional Neural Networks (CNN) to restore the color of grayscale images, and use the learned dictionary to convert grayscale images into color images.

4. **Video processing applications**:
- Apply ONMF to video frame processing, demonstrating its advantages in capturing time - evolving patterns in videos. In particular, compared with offline NMF, ONMF can better capture dynamic changes in videos.

### Formula Summary

- **Nonnegative Matrix Factorization (NMF) Optimization Problem**:
\[
\inf_{W \in \mathbb{R}^{d \times r}_{\geq 0}, H \in \mathbb{R}^{r \times n}_{\geq 0}} \|X - WH\|_F^2 + \lambda \|H\|_1
\]
where \(\|A\|_F\) represents the Frobenius norm of matrix \(A\), and \(\lambda \geq 0\) is the L1 regularization parameter.

- **Online Nonnegative Matrix Factorization (ONMF) Update Formulas**:
\[
H_t = \arg \min_{H \geq 0} \|X_t - W_{t - 1}H\|_F^2 + \lambda \|H\|_1
\]
\[
A_t=\frac{1}{t}((t - 1)A_{t - 1}+H_tH_t^T)
\]
\[
B_t=\frac{1}{t}((t - 1)B_{t - 1}+H_tX_t^T)
\]
\[
W_t = \arg \min_{W \geq 0}\frac{1}{2}\text{tr}(WA_tW^T)-\text{tr}(B_tW)
\]

Through these methods, the paper demonstrates the advantages of ONMF in processing large - scale, dynamic data sets, especially the application potential in image and time - series data processing.