Abstract
Multi-view learning tasks typically seek an aggregate synthesis of multiple views or perspectives of a single data set. The current approach assumes that there is an ambient space $X$ in which the views are images of $X$ under certain functions and attempts to learn these functions via a neural network. Unfortunately, such an approach neglects to consider the geometry of the ambient space. Hierarchically hyperbolic spaces (HHSes) do, however, provide a natural multi-view arrangement of data; they provide geometric tools for the assembly of different views of a single data set into a coherent global space, a CAT(0) cube complex. In this work, we provide the first step toward theoretically justifiable methods for learning embeddings of multi-view data sets into CAT(0) cube complexes. We present an algorithm which, given a finite set of finite metric spaces (views) on a finite set of points (the objects), produces the key components of an HHS structure. From this structure, we can produce a CAT(0) cube complex that encodes the hyperbolic geometry in the data while simultaneously allowing for Euclidean features given by the detected relations among the views.
Introduction
In modern data science, multi-view learning is essential for aggregating diverse perspectives of the same underlying object—such as multi-angle photography or multilingual text—into a unified representation. Conventional methods often treat this synthesis as a black-box task, employing neural networks to map views into a latent ambient space. However, this paradigm frequently suffers from "geometric blindness," ignoring the intrinsic structure of the data space, and lacks interpretability regarding how different views relate to one another.
CubeRep addresses these challenges by framing multi-view synthesis through the lens of geometric group theory. By utilizing Hierarchically Hyperbolic Spaces (HHSes), the authors provide a geometric toolkit to assemble disparate views into a global, interpretable structure known as a CAT(0) cube complex. This approach acknowledges that real-world data often exhibits a mixture of hierarchical (hyperbolic) and grid-like (Euclidean) geometries.
Methods
The core contribution of CubeRep is an algorithm designed to learn the "assembly instructions" for multi-view data. It identifies fundamental geometric relations between views $V$ and $W$:
- Orthogonal ($\perp$): The views are independent, analogous to orthogonal coordinates in a Euclidean plane.
- Transverse ($\pitchfork$): The views are related and impose constraints on each other.
Because the relations depend on discrete thresholds, the loss function's gradient is often zero, precluding standard gradient descent. Instead, the authors employ a Markov Chain Monte Carlo (MCMC) approach. This method performs a random walk over threshold values to optimize a loss function that penalizes cycles in learned relations, ensuring the output is a Directed Acyclic Graph (DAG) consistent with HHS theory requirements, such as Consistency Inequalities (CI) and Bounded Geodesic Image (BGI) properties.
Experimental Setup
The framework was evaluated using synthetic datasets where the ground truth geometric structure is known. These include simple graphs like the 4-cycle, representing independent views, and the "cross" graph, representing related views. More complex structures, such as trees, even cycles, and bipartite graphs without chords, were used to test the algorithm's ability to recover intricate mixtures of independent and transverse relations.
Results
Experimental results demonstrate that the MCMC-based algorithm robustly recovers the intended geometric structures. In the 4-cycle test, the model correctly identified view independence, while in the cross graph, it accurately detected that all views were related. On more complex bipartite and general graphs, the algorithm successfully distinguished between independent and transverse edges, effectively rebuilding the logic of the graph from metric data alone.
Consistency tests over 1,000 runs showed that the correct relations appeared persistently, indicating that the MCMC approach is reliable for these combinatorial optimization tasks. Unlike standard embeddings, CubeRep provides explicit, interpretable relations and a partial ordering of views.
Conclusion
CubeRep provides a theoretically grounded framework for learning the geometric relations underlying multi-view data. While currently limited to synthetic data and lacking support for "nested" relations, it marks a significant step toward interpretable, geometrically-aware representation learning. Future work will focus on automating the construction of the final CAT(0) cube complex, improving scalability for large-scale datasets, and validating the approach on high-dimensional real-world manifolds.
Summary updated