Abstract
We introduce coroICA, confounding-robust independent component analysis, a novel ICA algorithm which decomposes linearly mixed multivariate observations into independent components that are corrupted (and rendered dependent) by hidden group-wise stationary confounding. It extends the ordinary ICA model in a theoretically sound and explicit way to incorporate group-wise (or environment-wise) confounding. We show that our proposed general noise model allows to perform ICA in settings where other noisy ICA procedures fail. Additionally, it can be used for applications with grouped data by adjusting for different stationary noise within each group. Our proposed noise model has a natural relation to causality and we explain how it can be applied in the context of causal inference. In addition to our theoretical framework, we provide an efficient estimation procedure and prove identifiability of the unmixing matrix under mild assumptions. Finally, we illustrate the performance and robustness of our method on simulated data, provide audible and visual examples, and demonstrate the applicability to real-world scenarios by experiments on publicly available Antarctic ice core data as well as two EEG data sets.
Introduction
Independent Component Analysis (ICA) is a cornerstone of signal processing, traditionally used to solve the "cocktail party problem" by separating individual signal sources from a linear mixture. Standard ICA assumes the model $X = AS$, where $X$ is the observed data, $A$ is the mixing matrix, and $S$ are independent sources. However, real-world datasets are often plagued by confounding noise ($H$) that affects multiple sensors simultaneously, creating artificial dependencies that violate the core assumptions of standard ICA.
While traditional "noisy ICA" methods typically assume simple i.i.d. Gaussian noise, many scientific domains—ranging from climate science to neuroimaging—exhibit complex, structured noise. In these cases, standard algorithms like fastICA or SOBI often fail to recover the true underlying signals. This paper introduces coroICA, a framework designed to be robust against such confounding by leveraging group-wise stationarity.
Methodology: Confounding-Robust ICA
The coroICA model explicitly incorporates confounding noise into the mixing equation: $$X_i = A \cdot S_i + H_i$$ where $A$ is the invariant mixing matrix, $S_i$ are the independent sources, and $H_i$ represents the confounding noise for a specific group $i$ (e.g., a subject or a time block).
Group-Wise Stationary Confounding
The core innovation of coroICA is the assumption that while the signal $S$ is non-stationary (its variance or auto-covariance changes over time), the covariance of the noise $H$ remains constant within a defined group. This allows for a "subtraction trick": by taking the difference between covariance matrices at different time points within the same group, the stationary noise component cancels out: $$\text{Cov}(X_{\text{time1}}) - \text{Cov}(X_{\text{time2}}) = A [\text{Cov}(S_{\text{time1}}) - \text{Cov}(S_{\text{time2}})] A^\top$$
By jointly diagonalizing these noise-free difference matrices across multiple groups, coroICA can identify the mixing matrix $A$ even in the presence of heavy, structured confounding.
As illustrated in the comparison above, coroICA extends the capabilities of existing methods by allowing for group-wise stationary noise, whereas previous approaches like choiICA or SOBI are limited to time-independent noise or no noise at all.
Experimental Results
The authors evaluated coroICA across diverse datasets to demonstrate its robustness and utility:
- Simulated Data: In tests using GARCH models and shifting variance, coroICA maintained high accuracy (low Minimum Distance index) as noise levels increased, while competitors like fastICA and SOBI failed.
- Antarctic Ice Core Data: coroICA was applied to 800,000 years of CO2 and temperature data. It produced Equilibrium Climate Sensitivity (ECS) estimates more consistent with IPCC "likely" ranges than standard ICA, which was highly sensitive to parameter tuning.
- EEG Analysis: In multi-subject EEG studies, coroICA recovered source components (such as alpha waves and eye-blink artifacts) that were significantly more stable across subjects. Furthermore, using coroICA-recovered sources as features improved classification accuracy in Brain-Computer Interface (BCI) tasks compared to traditional methods.
Discussion and Conclusion
coroICA provides a theoretically sound extension to ICA that is particularly effective for grouped data. A significant practical advantage is that coroICA performs as well as standard ICA when no noise is present, making it a robust default choice for researchers.
However, the method requires that the signal itself be non-stationary; if the signal's variance and auto-covariance are perfectly constant, the model becomes unidentifiable. Additionally, the user must be able to define appropriate groups where noise is expected to be stationary. Despite these requirements, coroICA represents a significant step forward in making ICA applicable to complex, real-world confounding scenarios.
Software Availability: The authors have released a scikit-learn compatible Python package (coroICA), along with R and Matlab implementations, available at https://sweichwald.de/coroICA/.
Summary updated