Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Galaxy Morphology Classification with Counterfactual Explanation

Zhuo Cao, Lena Krieger, Hanno Scharr, Ira Assent

TL;DR

The paper addresses the interpretability gap in galaxy morphology classification by introducing an encoder–decoder augmented with an invertible flow to generate realistic counterfactual explanations. By splitting the latent space into class-dependent and class-independent components and enforcing an information bottleneck with an MMD-VAE objective, the approach learns compact, interpolable representations that enable controllable, minimal modifications across morphology classes. On Galaxy10 DECaLS with Galaxy Zoo 2 labels, the method achieves about 80% accuracy and produces high-quality counterfactuals (MSE around 0.006, SSIM around 0.96), while latent-space analyses reveal meaningful class separations and areas for data-quality assessment. The framework provides interpretable insights into which morphological features drive classifications and offers a practical tool for diagnosing model decisions and potential labeling issues in astronomical surveys.

Abstract

Galaxy morphologies play an essential role in the study of the evolution of galaxies. The determination of morphologies is laborious for a large amount of data giving rise to machine learning-based approaches. Unfortunately, most of these approaches offer no insight into how the model works and make the results difficult to understand and explain. We here propose to extend a classical encoder-decoder architecture with invertible flow, allowing us to not only obtain a good predictive performance but also provide additional information about the decision process with counterfactual explanations.

Galaxy Morphology Classification with Counterfactual Explanation

TL;DR

The paper addresses the interpretability gap in galaxy morphology classification by introducing an encoder–decoder augmented with an invertible flow to generate realistic counterfactual explanations. By splitting the latent space into class-dependent and class-independent components and enforcing an information bottleneck with an MMD-VAE objective, the approach learns compact, interpolable representations that enable controllable, minimal modifications across morphology classes. On Galaxy10 DECaLS with Galaxy Zoo 2 labels, the method achieves about 80% accuracy and produces high-quality counterfactuals (MSE around 0.006, SSIM around 0.96), while latent-space analyses reveal meaningful class separations and areas for data-quality assessment. The framework provides interpretable insights into which morphological features drive classifications and offers a practical tool for diagnosing model decisions and potential labeling issues in astronomical surveys.

Abstract

Galaxy morphologies play an essential role in the study of the evolution of galaxies. The determination of morphologies is laborious for a large amount of data giving rise to machine learning-based approaches. Unfortunately, most of these approaches offer no insight into how the model works and make the results difficult to understand and explain. We here propose to extend a classical encoder-decoder architecture with invertible flow, allowing us to not only obtain a good predictive performance but also provide additional information about the decision process with counterfactual explanations.
Paper Structure (17 sections, 4 equations, 7 figures, 2 tables)

This paper contains 17 sections, 4 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Data and Methodology
  3. Visual counterfactual explanations
  4. Model architecture
  5. Invertible Flow
  6. Loss Functions
  7. Galaxy10 DECaLS
  8. Results
  9. Metrics
  10. Counterfactual Explanation
  11. Latent Space Analysis
  12. Discussion
  13. Conclusion
  14. Appendix / Supplemental Material
  15. Experiment setting
  16. ...and 2 more sections

Figures (7)

  • Figure 1: Architecture of our model. Left panel: training phase. Right panel: explanation phase.
  • Figure 2: Sample images for each class of Galaxy 10 DECaLS and number of instances.
  • Figure 3: Original images aligned across diagonal (red boxes) with counterfactuals displayed in the same row classified according to column labels. Due to limited space, a complete image grid is shown in Appendix \ref{['app: complete_visualization']}.
  • Figure 4: t-SNE plots for class-dependent ($\mathbf{z}_1$), hidden ($\mathbf{h}_1$) and background ($\mathbf{z}_2$) features.
  • Figure 5: Confusion matrix of the trained model for the entire dataset.
  • ...and 2 more figures