GSVD for Geometry-Grounded Dataset Comparison: An Alignment Angle Is All You Need

This paper proposes a geometry-grounded framework for comparing datasets using the Generalized Singular Value Decomposition (GSVD) to derive an interpretable "angle score" that quantifies whether individual samples are better explained by one dataset, the other, or both.

The Gist

Here is an explanation of the paper "GSVD for Geometry-Grounded Dataset Comparison," translated into simple language with creative analogies.

The Big Idea: Comparing Datasets Without Losing the Plot

Imagine you have two huge libraries of books.

• Library A is filled with mystery novels.
• Library B is filled with sci-fi novels.

Usually, if you want to compare them, you might ask a librarian (a complex AI model) to read a few pages and tell you which library a new book belongs to. But this paper asks a different question: Can we compare the libraries themselves by looking at their "architecture" or "geometry"?

The authors propose a new way to look at data that treats every piece of information not just as a list of numbers, but as a direction in space. They want to know: Does this new book feel more like it belongs in the Mystery wing or the Sci-Fi wing?

The Problem: "Arbitrary Vectors" vs. "Geometry"

Most AI today treats data like a bag of random ingredients. It doesn't care that "red" and "blue" are related colors, or that "running" and "walking" are related actions. It just sees numbers.

This paper says: Stop treating data like a bag of marbles. Start treating it like a map.
If you have a map of the world, you can see that Paris and London are close, while Tokyo is far away. The authors want to build a map for their datasets so they can measure the "distance" and "angle" between them.

The Solution: The "Universal Translator" (GSVD)

To compare the two libraries (datasets), the authors use a mathematical tool called GSVD (Generalized Singular Value Decomposition).

The Analogy: The Shared Dance Floor
Imagine two groups of dancers: Group A (Mystery fans) and Group B (Sci-Fi fans). They are dancing in a huge room.

• Sometimes, they dance in a way that is unique to them (Mystery fans do a specific spin; Sci-Fi fans do a specific jump).
• Sometimes, they dance in a way that is the same (both groups clap their hands).

The GSVD is like a magic camera that finds a "Shared Dance Floor" (a common coordinate system). It separates the moves into three categories:

1. The Mystery Moves: Unique to Group A.
2. The Sci-Fi Moves: Unique to Group B.
3. The Shared Moves: Moves both groups do.

This camera creates a "Joint Frame of Reference." Now, instead of looking at the messy original room, everyone is viewed through this clean, shared lens.

The Star of the Show: The "Alignment Angle" ($\theta$)

Once the data is on this shared dance floor, the authors introduce a simple score called the Alignment Angle. Think of this as a compass for a new piece of data (a new book, or a new image).

When a new item arrives, the compass points in a direction. The angle tells you everything you need to know:

• Angle near 0°: The item is purely Mystery. It fits perfectly with Group A's unique moves.
• Angle near 90°: The item is purely Sci-Fi. It fits perfectly with Group B's unique moves.
• Angle near 45°: The item is ambiguous. It's doing a mix of both, or it's a "Shared Move" that fits neither group perfectly. It's like a book that is a "Sci-Fi Mystery."

Why is this cool?
Instead of a black-box AI saying "I'm 85% sure this is a Mystery," this method gives you a geometric reason: "This book is at a 10-degree angle from the Mystery direction, so it's definitely a Mystery." It's transparent and easy to understand.

How They Tested It: The MNIST Experiment

The authors tested this on MNIST, a famous dataset of handwritten digits (0 through 9).

• They built a "Mystery Library" out of images of the number 4.
• They built a "Sci-Fi Library" out of images of the number 9.

The Results:

1. Clear Separation: When they tested images of 4s, the compass pointed almost straight to 0°. When they tested 9s, it pointed to 90°.
2. The "Fuzzy" Ones: When they looked at the number 4 vs. 9, they found that some 4s looked a bit like 9s (maybe a curly tail). The compass for those specific images pointed to 45°.
3. Visualizing the "Extreme" Directions: They could even generate "ghost images" of what a perfect 4 looks like according to their math, and what a perfect 9 looks like. These ghost images showed exactly why the computer thought they were different (e.g., the sharp angles of the 4 vs. the round loops of the 9).

Why Does This Matter?

1. No More Black Boxes: Instead of guessing why an AI made a mistake, you can look at the angle and say, "Ah, this image was at 45 degrees, so the AI was confused because it looked like both classes."
2. Better Data Cleaning: If you have a dataset full of "bad" data (like a photo of a cat labeled as a dog), this angle will be weird. It will point somewhere in the middle, flagging it for a human to check.
3. Understanding Similarity: It helps us understand how two things are similar. Are they similar because they share a lot of features, or because they are just both "vague"?

Summary in One Sentence

This paper gives us a geometric compass that measures exactly how much a piece of data belongs to one group versus another, turning complex math into a simple angle that tells us if something is "Team A," "Team B," or "Confused."

Technical Summary

Here is a detailed technical summary of the paper "GSVD FOR GEOMETRY-GROUNDED DATASET COMPARISON: AN ALIGNMENT ANGLE IS ALL YOU NEED."

1. Problem Statement

The paper addresses the challenge of comparing two datasets (e.g., two classes of images, or data from two different domains) in a way that respects their underlying geometric structure.

• Limitations of Current Methods: Existing approaches often compare datasets indirectly via trained models, embedding distances, or distributional metrics (like MMD or FID). These methods can obscure why datasets are similar or different, often treating observations as arbitrary vectors rather than points in a structured geometric space.
• The Goal: To establish a "geometry-grounded" primitive for dataset comparison that identifies shared latent structures versus dataset-specific features without requiring point-to-point sample correspondences (which are often unavailable or unreliable).

2. Methodology

The proposed framework relies on Generalized Singular Value Decomposition (GSVD) to create a joint coordinate system for two datasets.

A. The Co-Span Relation Primitive

Instead of mapping samples directly, the authors define similarity through a linear relation in a shared ambient space. Given two dataset matrices $A \in \mathbb{R}^{d \times p}$ and $B \in \mathbb{R}^{d \times q}$ (where columns are observations), they seek vectors $x$ and $y$ such that:
$$Ax = By = z$$
Here, $z$ is a shared ambient vector. This formulation encodes compatibility between the column spaces of $A$ and $B$ without requiring an invertible mapping between the datasets.

B. GSVD as a Joint Coordinate System

The authors utilize GSVD to decompose the pair $(A, B)$ into a shared reference frame:
$$A = HCU, \quad B = HSV$$
Where:

• $H \in \mathbb{R}^{d \times d}$ is an invertible (or left-invertible) matrix defining a shared ambient reference frame.
• $U, V$ are orthogonal matrices.
• $C$ and $S$ are diagonal (or block-diagonal) matrices with non-negative entries satisfying $C^\top C + S^\top S = I$.

Interpretation of $(C, S)$:
The diagonal entries of $C$ and $S$ quantify the contribution of each shared direction in $H$ to dataset $A$ versus dataset $B$.

• High $C_{ii}$, low $S_{ii}$: Direction is specific to $A$.
• Low $C_{ii}$, high $S_{ii}$: Direction is specific to $B$.
• Comparable magnitudes: Direction represents shared structure.

C. The Alignment Angle $\theta(z)$

The core contribution is a per-sample diagnostic score, the Alignment Angle $\theta(z) \in [0, \pi/2]$, derived from the GSVD factors.

1. Computation: For a sample $z$, project it into the shared frame: $c(z) = H^\dagger z$.
2. Cost Calculation: Compute the "cost" to represent $z$ using $A$ and $B$ respectively:
   • $a(z) = \|C^\dagger c(z)\|_2$
   • $b(z) = \|S^\dagger c(z)\|_2$
3. Angle Definition:
   $$\theta(z) = \arctan\left(\frac{a(z)}{b(z)}\right)$$

Interpretation of $\theta(z)$:

• $\theta(z) \approx 0$: The sample is explained more economically by dataset $A$ ("More A").
• $\theta(z) \approx \pi/2$: The sample is explained more economically by dataset $B$ ("More B").
• $\theta(z) \approx \pi/4$: The sample lies in the shared structure, explained comparably by both.

D. Extreme Directions

The paper also derives the specific vectors $z_{max}$ and $z_{min}$ that maximize and minimize $\theta(z)$. These correspond to specific columns of the shared matrix $H$ (specifically $h_{r+k}$ and $h_{r+1}$), providing visualizable "prototypical" directions for each dataset and their intersection.

3. Key Contributions

1. Co-Span Primitive: Proposes the linear relation $Ax=By=z$ as a minimal, geometry-grounded primitive for comparing datasets without requiring sample correspondences.
2. GSVD Joint Frame: Utilizes GSVD to explicitly separate shared versus dataset-specific directions via the diagonal factors $(C, S)$, creating an interpretable coordinate system.
3. Alignment Angle Score: Derives $\theta(z)$, a scalar, interpretable metric that quantifies the relative explanatory power of two datasets for a specific sample.
4. Geometric Diagnostics: Demonstrates that the GSVD frame yields representative directions (extremes) that visualize what is unique to a class and what is shared.
5. Information-Geometric Connection: Links the linear algebraic angle $\theta$ to probabilistic concepts, showing that differences in $\theta$ correspond to Fisher-Rao distances between induced Bernoulli posteriors.

4. Experimental Results

The authors evaluated the method on the MNIST dataset (digits) and Fashion-MNIST.

• Setup: Two classes (e.g., digit "1" vs. "5") were used to construct matrices $A$ and $B$. The alignment angle $\theta(z)$ was computed for all test samples.
• Histogram Analysis:
  • Distinct Classes (e.g., "1" vs. "5"): The distributions of $\theta(z)$ for the two classes were well-separated. Class "1" samples clustered near 0, and Class "5" samples clustered near $\pi/2$.
  • Similar Classes (e.g., "4" vs. "9"): The distributions showed significant overlap near $\pi/4$, reflecting the visual and geometric ambiguity between these digits.
• Extreme Directions Visualization: The authors reconstructed the "extreme" vectors ($z_{max}$ and $z_{min}$) as images.
  • The "4-aligned" extreme showed sharp, edge-like patterns typical of a 4.
  • The "9-aligned" extreme showed rounded contours.
  • The "shared" direction showed a blended structure containing features of both.
• Fisher-Rao Distance: The paper quantified the separation between class-conditional $\theta$-histograms using Fisher-Rao distance. Pairs with high visual similarity (like 4 vs. 9) had lower Fisher-Rao distances, while distinct pairs had higher distances.
• Classification: A simple binary classifier using a threshold $\tau = \pi/4$ on $\theta(z)$ was presented. While not intended as a state-of-the-art classifier, it demonstrated the score's utility as a diagnostic tool for identifying samples that align with one domain over another.

5. Significance and Implications

• Interpretability: Unlike "black box" distance metrics, this method provides a geometric explanation for dataset similarity. It explicitly identifies which directions are shared and which are unique.
• No Correspondence Required: The method works effectively even when there is no one-to-one mapping between samples in the two datasets, making it suitable for comparing different domains or classes.
• Diagnostic Utility: The angle score serves as a powerful tool for auditing datasets, identifying outliers, or filtering samples during transfer learning (e.g., downweighting samples that do not align with the target domain).
• Theoretical Bridge: The work connects linear algebra (GSVD) with information geometry (Fisher-Rao distance), offering a unified view of dataset comparison.
• Future Directions: The authors suggest extending the framework to multiple domains (multi-way GSVD) and applying it to learned feature embeddings (e.g., from Transformers), where linear subspace assumptions may be even more valid due to the linearization of semantic structure.

In summary, the paper proposes a mathematically rigorous and highly interpretable framework for dataset comparison, shifting the focus from distribution matching to geometric alignment via the Generalized Singular Value Decomposition.