Robust Sparse Blind Source Separation

Cecile Chenot, Jérôme Bobin, Jérémy Rapin

Abstract: Blind Source Separation is a widely used technique to analyze multichannel data. In many real-world applications, its results can be significantly hampered by the presence of unknown outliers. In this paper, a novel algorithm coined rGMCA (robust Generalized Morphological Component Analysis) is introduced to retrieve sparse sources in the presence of outliers. It explicitly estimates the sources, the mixing matrix, and the outliers. It also takes advantage of the estimation of the outliers to further implement a weighting scheme, which provides a highly robust separation procedure. Numerical experiments demonstrate the efficiency of rGMCA to estimate the mixing matrix in comparison with standard BSS techniques.

Introduction

Blind Source Separation (BSS) is a fundamental problem in data science, often described as the "cocktail party problem": given several mixed signals, the goal is to recover the original individual sources. While standard methods like Independent Component Analysis (ICA) or sparse BSS perform well under ideal conditions, they are notoriously fragile in the presence of outliers. In fields such as astrophysics or medical imaging, data is frequently corrupted by spurious spikes, sensor glitches, or impulsive noise. These outliers do not follow a standard Gaussian distribution, and even a single massive outlier can significantly bias the estimation of the mixing matrix, leading to poor source recovery. Existing solutions often require risky pre-processing or are limited to specific data types, such as non-negative signals.

Methodology

The proposed algorithm, rGMCA (robust Generalized Morphological Component Analysis), builds upon the standard GMCA framework which leverages signal sparsity for separation. The core innovation of rGMCA lies in its two-layered defense against outliers:

  1. Joint Estimation: Rather than cleaning data as a pre-processing step, rGMCA explicitly models the observed data $X$ as a combination of mixed sources, outliers, and noise: $X = AS + O + N$. It jointly estimates the mixing matrix $A$, the sources $S$, and the outliers $O$.
  2. Iterative Weighting Scheme: To robustly estimate the mixing matrix, rGMCA employs a weighting matrix $W$. This scheme assigns lower weights to data samples identified as likely outliers, effectively "ignoring" corrupted entries during the update of the mixing matrix.

The algorithm utilizes a "decreasing threshold" strategy, initially focusing on the most distinct signal components and iteratively refining the estimates as noise and outliers are removed.

Experimental Setup

The performance of rGMCA was evaluated using both synthetic and realistic datasets:

  • Simulated Synthetic Data: Mixtures of 8 sources across 16 observations were generated using Bernoulli-Gaussian distributions. Outliers were introduced as corrupted columns and random spikes to test robustness across various patterns.
  • NMR Spectroscopy Data: Simulated Nuclear Magnetic Resonance (NMR) spectra for four molecules (menthone, folic acid, ascorbic acid, and myo-inositol) provided a realistic biomedical test case for identifying chemical signatures in complex mixtures.

Key Results and Discussion

Numerical experiments highlight the superior robustness of rGMCA compared to standard BSS techniques. A critical finding is that large outliers violate the Morphological Diversity Property (MDP)—the assumption that sources do not peak simultaneously. Standard algorithms are often "tricked" into treating these dominant outlier peaks as sources.

Figure 1: Mixing matrix error (Δ_A) as a function of outlier amplitude. rGMCA maintains high accuracy even when outliers are four orders of magnitude larger than the signal.

As shown in the figure above, rGMCA maintains a consistently low error rate (Δ_A) even as the amplitude of outliers reaches $10^4$ times the signal level, whereas competing methods like standard GMCA and β-divergence fail rapidly. Furthermore, rGMCA demonstrates a high breakdown point, handling up to 10% data corruption with near-perfect success rates.

Figure 2: Mixing matrix error (Δ_A) versus the number of observations. rGMCA achieves accurate separation with significantly fewer observations than other robust methods.

Another significant advantage of rGMCA is its efficiency regarding sample size. While methods based on Principal Component Pursuit (PCP) typically require a large number of observations ($m \gg n$), rGMCA remains effective even when the number of observations is close to the number of sources. In NMR tests with 4 sources, rGMCA provided correct estimates with as few as 5 observations.

Conclusion

rGMCA provides a robust and efficient framework for sparse Blind Source Separation in the presence of significant outliers. By jointly estimating outliers and employing an iterative weighting scheme, it overcomes the limitations of traditional BSS methods. While the current implementation excels at mixing matrix estimation, future work will focus on improving the separation of outliers from sources in the final output and extending the method to transformed domains like wavelets to handle non-sparse sources.

Summary updated

Abstract

Blind Source Separation is a widely used technique to analyze multichannel data. In many real-world applications, its results can be significantly hampered by the presence of unknown outliers. In this paper, a novel algorithm coined rGMCA (robust Generalized Morphological Component Analysis) is introduced to retrieve sparse sources in the presence of outliers. It explicitly estimates the sources, the mixing matrix, and the outliers. It also takes advantage of the estimation of the outliers to further implement a weighting scheme, which provides a highly robust separation procedure. Numerical experiments demonstrate the efficiency of rGMCA to estimate the mixing matrix in comparison with standard BSS techniques.
BibTeX
@article{chenot2015robust,
  author = {Cecile Chenot and Jérôme Bobin and Jérémy Rapin},
  title = {{Robust Sparse Blind Source Separation}},
  journal = {IEEE Signal Processing Letters},
  year = {2015},
  url = {https://openreview.net/forum?id=ycfNKfhzAd},
}
Figure