A Mean Curvature Approach to Boundary Detection: Geometric Insights for Unsupervised Learning

Technical Summary: Mean Curvature Boundary Points (MCBP)

Problem Statement
Precise detection of boundaries in high-dimensional, unsupervised learning remains a significant challenge, particularly when data exhibit non-linear structures, anisotropic distributions, or heterogeneous densities. Traditional approaches rely predominantly on density-based principles (e.g., DBSCAN, LOF) and assume that boundary points lie in low-density regions. However, the present article argues that density variations alone are insufficient to characterize complex data manifolds. Purely density-driven methods frequently fail to distinguish flat low-density regions from genuinely high-curvature transition zones (concave or convex interfaces), leading to imprecise boundary detection in scenarios involving non-linear manifolds and the "curse of dimensionality."

Methodology: Mean Curvature Boundary Points (MCBP)
The proposed MCBP framework departs from density estimation by explicitly modeling the intrinsic geometry of the data manifold using differential geometry. The fundamental assumption is that boundary points correspond to regions of high mean curvature, capturing geometric transitions overlooked by density-based methods.

The algorithm operates through the following computational pipeline:

1. Construction of Local Neighborhoods: For each sample $x_i$, a $k$-nearest-neighbor ($k$-NN) patch $P_i$ is constructed.
2. Approximation of the Metric Tensor (First Fundamental Form): The local metric structure is approximated using the inverse of the empirical covariance matrix ($\Sigma_i^{-1}$) of the neighborhood. This serves as a discrete analogue to the metric tensor and captures local anisotropy and feature dependencies.
3. Approximation of the Second Fundamental Form: A local Hessian operator is estimated by fitting a second-degree polynomial (quadratic terms) to the local neighborhood. The resulting matrix ($H_i$) approximates the Second Fundamental Form and encodes how the manifold curves.
4. Shape Operator and Mean Curvature: The discrete shape operator $S_i$ is computed as the product of the Hessian approximation and the covariance matrix ($S_i \approx H_i \Sigma_i$). The mean curvature $K_i$ for each point is defined as the trace of this operator ($K_i = \text{tr}(S_i)$).
5. Adaptive Thresholding: Boundary points are identified using an adaptive, percentile-based threshold. Points with curvature values above a specific percentile (e.g., the top 25%) are labeled as boundary points. This mechanism enables multi-scale boundary extraction without dependence on arbitrary density parameters.

Main Contributions
The article outlines three primary contributions:

1. Geometric Criterion for Boundary Detection: MCBP establishes mean curvature as a principled replacement for boundary characterization. It unifies the identification of outliers, transition regions, and decision boundaries under a single mathematical framework, moving beyond density assumptions to capture intrinsic geometric complexity.
2. Adaptive Multi-scale Extraction: The introduction of a percentile-based threshold mechanism enables flexible control over boundary thickness and allows the method to adapt to varying data densities without requiring sensitive parameter tuning.
3. Curvature-Driven Data Decomposition: The method proposes a strategy to partition datasets into "smooth" (low curvature) and "boundary" (high curvature) subsets. This acts as a non-linear geometric filter, effectively separating stable core regions from geometrically ambiguous or noisy transition zones.

Experimental Results
The authors evaluated MCBP on 25 real-world high-dimensional datasets from the OpenML repository, as well as several synthetic 2D datasets (Gaussian blobs, anisotropic ellipses, and two moons).

• Synthetic Data: Visualizations confirmed that MCBP correctly identifies outer contours and transition regions between clusters, even in anisotropic and non-convex settings where density-based methods struggle.
• Clustering Performance (Preprocessing): When used as a preprocessing step to filter high-curvature points prior to applying $k$-means, the method consistently improved clustering quality. Across 25 datasets, the average Silhouette coefficient increased from 0.193 to 0.288, and the Calinski–Harabasz index improved significantly, indicating better cluster compactness and separation.
• Centroid Initialization: Using centroids derived from the filtered (smooth) dataset to initialize $k$-means on the original data yielded more stable results than standard $k$-means++ initialization, particularly in noisy or complex geometries.
• Density-Based Clustering (HDBSCAN): Applying MCBP filtering prior to HDBSCAN significantly improved performance, with average Silhouette coefficients rising from 0.227 to 0.359. This suggests that removing high-curvature noise stabilizes density estimation.
• Hybrid Strategies: A hybrid approach using HDBSCAN on the smooth subset and 1-NN for boundary propagation showed substantial gains, with average Silhouette coefficients more than doubling (from 0.1047 to 0.2308) and the Calinski–Harabasz index tripling.

Significance and Claims
The article positions MCBP as a concrete contribution to Geometric Machine Learning, bridging differential geometry and data-driven modeling. The authors claim that:

• Curvature as a Unifying Paradigm: Mean curvature offers a unified geometric interpretation of boundaries, outliers, and transitions, providing a more robust characterization than density alone, especially in high-dimensional and non-linear settings.
• Principled Preprocessing: Curvature-driven decomposition is not merely a diagnostic tool but an active component for improving downstream unsupervised learning procedures. By filtering geometrically unstable regions, the method enhances the recoverability of intrinsic cluster structures.
• Robustness: The approach is less susceptible to global distribution distortions and the "empty space phenomenon" common in high dimensions, as it relies on local geometric operators rather than global density estimates.

The work concludes that incorporating differential geometric priors, particularly mean curvature, provides a principled mechanism to improve the stability, separability, and interpretability of unsupervised learning pipelines.