Riemannian Archetypal Analysis: Interpretable non-linear data analysis on deformed star distributions

Background

Classical archetypal analysis (AA) is valued for its interpretability, as it represents data points as mixtures of extremal "archetypes." However, its reliance on linear convex combinations limits its ability to model complex, non-linear data structures. Existing neural extensions to AA can capture non-linearity but often compromise the geometric interpretability and meaningfulness of the archetypes and interpolations that is central to the classical method's appeal. The core challenge is to develop a method that retains geometric interpretability while gaining the expressive power needed for non-linear data.

Key Points

• Core Innovation: Data-Driven Pullback Geometry. The paper develops a Riemannian AA based on pullback geometry derived from the data itself, rather than assuming a predefined metric space. This allows the model to adapt its underlying geometry to the data's intrinsic structure.
• Theoretical Foundation: Deformed Star Distributions. The authors introduce a class of statistical models called deformed star distributions. These distributions, along with their associated pullback Riemannian geometry, provide a statistical interpretation for the non-linear mappings learned by the model.
• Model Definition: Riemannian Archetypal Mapping (RAM). The RAM is formally defined as a projection onto the manifold formed by geodesically convex combinations of the archetypes. This generalizes the classical linear convex hull to a geodesically convex hull on the learned data manifold.
• Optimization Strategy. A practical training scheme is proposed that first finds a convex relaxation of the problem for initialization, followed by a non-convex refinement stage. This two-phase approach addresses the computational complexity of optimizing the manifold mapping and archetype combinations.
• Learning Framework. The work includes a method to learn the parameters of the deformed star distributions directly from data. The authors note this yields "reasonable, albeit generally suboptimal" distributions, indicating an area for further improvement.
• Experimental Validation. Tests on synthetic data and the MNIST dataset show the framework produces:
  • Meaningful geodesics between archetypes that respect the data manifold.
  • Useful projections for tasks like denoising.
  • Geometry-aware classifications.
  • The experiments also serve to clarify where current optimization limitations remain, providing transparency about the method's current stage of development.

Significance

This research offers a principled bridge between two powerful but historically disparate paradigms: the geometrically interpretable classical AA and the expressive non-linear neural models. The significance lies in its attempt to formally reintroduce geometric meaning—through concepts like geodesic convexity and pullback metrics—into non-linear archetypal analysis. By doing so, it aims to recover the core interpretive benefit of archetypes (understanding data as mixes of extremes) while operating on complex data manifolds. The explicit identification of optimization challenges also provides valuable direction for future work in this hybrid domain.