The Manifold Density Function: An Intrinsic Method for the Validation of Manifold Learning

Benjamin Holmgren,Eli Quist,Jordan Schupbach,Brittany Terese Fasy,Bastian Rieck

2024-02-15

Abstract:We introduce the manifold density function, which is an intrinsic method to validate manifold learning techniques. Our approach adapts and extends Ripley's $K$-function, and categorizes in an unsupervised setting the extent to which an output of a manifold learning algorithm captures the structure of a latent manifold. Our manifold density function generalizes to broad classes of Riemannian manifolds. In particular, we extend the manifold density function to general two-manifolds using the Gauss-Bonnet theorem, and demonstrate that the manifold density function for hypersurfaces is well approximated using the first Laplacian eigenvalue. We prove desirable convergence and robustness properties.

Machine Learning,Algebraic Topology

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to verify the effectiveness of manifold learning algorithms in an unsupervised setting. Specifically, the author proposes a new method - the manifold density function - to evaluate whether manifold learning algorithms can effectively capture the latent manifold structure in data. This method does not rely on any prior knowledge about the true geodesic distance, but judges the performance of the algorithm by evaluating whether the data points output by the algorithm are locally similar to samples uniformly distributed in Euclidean space.

### Main Contributions
1. **Introduction of the Manifold Density Function**:
- A new density estimator \(K_X\) is defined. This function has good convergence properties and can be accurately calculated when the scalar curvature is known.
- The formula is as follows:
\[
K_X(r) := \frac{\text{Vol}(B_2(0, r))}{\text{Vol}(X)}
\]
where \(\text{Vol}(B_2(0, r))\) is the volume of the Euclidean ball with radius \(r\), and \(\text{Vol}(X)\) is the volume of the manifold \(X\).

2. **Robust Approximation on Two - Dimensional Manifolds**:
- An approximation method that can be efficiently computed on two - dimensional manifolds is proposed. It is estimated using the Euler characteristic number and has provable accuracy.
- The approximation formula is:
\[
\hat{K}_p(r) := \left(1 - \frac{S(p)\cdot r^2}{6(n + 2)}\right)^{-1}\cdot\frac{1}{|X|}\sum_{x\in X}I(x\in B_r(p))
\]
where \(S(p)\) is the scalar curvature at point \(p\), and \(I(x\in B_r(p))\) is an indicator function indicating whether point \(x\) is within the ball with radius \(r\) centered at \(p\).

3. **Robust Approximation on High - Dimensional Hypersurfaces**:
- An approximation method that can be efficiently computed on high - dimensional hypersurfaces is proposed. It is estimated using the first Laplace eigenvalue and has provable accuracy.
- The approximation formula is:
\[
\hat{K}(r)\approx\left(1 - \frac{r^2\cdot\lambda_1(n - 1)}{12(n)(n + 2)}\right)^{-1}\cdot\frac{1}{|X|^2}\sum_{p\in X}\sum_{x\in X}I(x\in B_r(p))
\]
where \(\lambda_1\) is the first eigenvalue of the Laplace operator.

### Core Ideas of the Method
- **Local Manifold Density Function**: Judge the performance of the manifold learning algorithm by evaluating whether the data points in each local neighborhood are similar to uniformly distributed samples.
- **Global Aggregation**: Aggregate the evaluation results of all local neighborhoods to obtain a global manifold density function, thereby providing an evaluation of the entire data set.

### Advantages
- **Intrinsic Verification**: Does not require any prior knowledge about the true geodesic distance and is completely verified based on the characteristics of the data itself.
- **Theoretical Guarantee**: Provides theoretical guarantees of convergence and robustness and is applicable to a wide range of Riemannian manifolds.
- **Efficient Computation**: Proposes efficient approximation methods that can be quickly computed in practical applications.

### Application Scenarios
- **Evaluation of Manifold Learning Algorithms**: Can be used to evaluate the performance of various manifold learning algorithms, especially in an unsupervised setting.
- **Evaluation of Data Uniformity**: Can be used to evaluate whether the data set is uniformly distributed on the manifold, which is an important prerequisite for many machine learning tasks.

In summary, this paper proposes a new, intrinsic method to verify the effectiveness of manifold learning algorithms, filling this gap in this field.