groupICA: Independent component analysis for grouped data

Niklas Pfister, Sebastian Weichwald, Peter Bühlmann, Bernhard Schölkopf

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. We provide a scikit-learn compatible pip-installable Python package coroICA as well as R and Matlab implementations accompanied by a documentation at https://sweichwald.de/coroICA/

Introduction

Independent Component Analysis (ICA) is a fundamental technique for "blind source separation," famously illustrated by the cocktail party problem where one seeks to isolate individual voices from a single recording of a crowded room. Standard ICA algorithms typically assume that data points are independent and identically distributed (i.i.d.). However, in many real-world applications, data is naturally grouped. For example, in an EEG study, recordings from multiple subjects share underlying brain "sources," but each subject introduces unique "confounders" such as different head shapes, electrode placements, or physiological noise.

The common "quick fix" of pooling all subjects' data into one giant matrix and running standard ICA is fundamentally flawed. Pooling ignores the group structure, causing algorithms to mix group-specific noise into the recovered signals. This leads to unstable components that fail to generalize across different subjects or environments. We introduce coroICA (confounding-robust ICA) to address this by explicitly modeling hidden group-wise confounding.

The coroICA Framework

The core of coroICA is a new statistical model: $X_i = A \cdot (S_i + H_i)$.

  • $X_i$: The observed multivariate data for group $i$.
  • $A$: A mixing matrix assumed to be fixed across all groups.
  • $S_i$: The independent source signals of interest.
  • $H_i$: Group-wise confounding variables.

Crucially, the covariance of the noise ($H$) is assumed to be constant within a group but can change between groups. This "group-wise stationary" assumption allows the algorithm to isolate signal changes. Because the noise has a fixed covariance within a group, taking the difference between covariance matrices of two different time segments within that same group causes the noise term to cancel out. What remains is a signal that depends only on the changes in the source signals ($S$).

Comparative Analysis

coroICA distinguishes itself from existing ICA methods by its flexibility regarding signal types and its unique noise tolerance. While traditional methods like fastICA rely on non-Gaussianity and allow no noise, coroICA leverages non-stationarity—the fact that signal variances change over time—and is robust to group-wise stationary noise.

Table 1: Comparison of coroICA with other ICA methods in terms of signal requirements and noise tolerance.

As shown in Table 1, coroICA is the only method among those compared that explicitly allows for group-wise stationary noise, making it uniquely suited for multi-subject or multi-environment datasets where other noisy ICA procedures fail.

Methodology and Results

The estimation procedure involves partitioning each group into smaller subgroups, computing covariance differences to isolate signal changes, and using an Approximate Joint Diagonalization algorithm (uwedge) to find a single unmixing matrix. Furthermore, coroICA ranks components based on stability across observations rather than just variance, identifying signals that are more likely to be biologically relevant.

Key Findings:

  • Robustness to Noise: In simulations, coroICA maintained high correlation accuracy with ground-truth sources even as confounding noise increased, whereas the performance of pooled ICA plummeted.
  • Better Generalization: When training on a subset of subjects and testing on new, unseen subjects, coroICA achieved significantly higher classification accuracy for Brain-Computer Interface (BCI) tasks compared to pooled ICA.
  • Biological Plausibility: Topographic maps produced by coroICA were "cleaner" and more closely resembled known dipolar patterns of brain activity, while pooled ICA often produced maps cluttered with noise artifacts.

Conclusion

By leaning on the idea that the true underlying physical mixing process ($A$) should be invariant across environments while noise ($H$) shifts, coroICA provides a powerful tool for analyzing grouped data. While it assumes a fixed mixing matrix and does not perfectly "denoise" the additive $H$ term, it successfully identifies stable, scientifically meaningful components. The authors have released coroICA as a scikit-learn compatible Python package to facilitate its use in neuroscience and signal processing research.

Summary updated

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. We provide a scikit-learn compatible pip-installable Python package coroICA as well as R and Matlab implementations accompanied by a documentation at https://sweichwald.de/coroICA/
BibTeX
@article{pfister2019groupica,
  author = {Niklas Pfister and Sebastian Weichwald and Peter Bühlmann and Bernhard Schölkopf},
  title = {{groupICA: Independent component analysis for grouped data}},
  journal = {J. Mach. Learn. Res.},
  year = {2019},
  url = {https://openreview.net/forum?id=fGECpbH1qI},
}
Figure