Robust Statistical Learning via Data Depth

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 112 papers16 conceptsfrozen 2026-10-04

In traditional statistics, we often describe a dataset using its mean and variance. However, these metrics are notoriously fragile: a single extreme outlier can pull the mean far away from the "bulk" of the data. Data depth offers a more resilient alternative. Instead of calculating averages, it assigns every point a value representing how "deep" or "central" it is within a data cloud.

By prioritizing geometric centrality over parametric assumptions, depth-based methods provide a robust framework for classification and anomaly detection that remains stable even when the training data is heavily contaminated.

The Intuition of Centrality

Imagine a cloud of points in high-dimensional space. A point on the very edge of this cloud has low depth, while a point buried in the thickest part of the cluster has high depth. This transformation—moving from raw coordinates to a "depth space"—allows us to analyze data based on its relative positioning.

As Oleksii Pokotylo, Rainer Dyckerhoff, and Stanislav Nagy demonstrate in their work on the ddalpha package, this approach is inherently non-parametric. It doesn't care if your data follows a Gaussian distribution or has heavy tails; it only cares about the underlying geometry. This makes it particularly effective for real-world datasets that violate the "clean" assumptions required by linear models.

Resilience to Contamination

The primary advantage of depth-based methods is their "breakdown point"—the amount of noise or outliers a model can handle before it fails. While standard loss-minimization models (like neural networks or linear regressions) can be skewed by a small percentage of malicious or erroneous data, depth functions are remarkably stubborn.

Research by Pavlo Mozharovskyi and Romain Valla in Anomaly detection using data depth: multivariate case shows that these measures can maintain stable detection boundaries even when up to 50% of the training data is contaminated. Because the "center" is defined by the collective geometry of the majority, a cluster of outliers on the periphery cannot easily shift the model's definition of what is "normal."

Measuring the Depth of a Curve

The concept of depth becomes even more powerful when applied to Functional Data Analysis (FDA), where each data point is not a single number but an entire curve (e.g., a temperature profile over a day or a stock price over a week).

Traditional methods for curves often rely on point-by-point averages, which can miss "shape anomalies"—curves that look normal at any single time point but have an unusual overall trajectory. To solve this, Guillaume Staerman, Pavlo Mozharovskyi, and Stephan CLEMENCON introduced the Area of the Convex Hull (ACH) Depth in The Area of the Convex Hull of Sampled Curves.

Instead of looking at points, ACH Depth looks at the global geometry of the curve. By measuring the area of the convex hull formed by sampled segments of the curve, the algorithm can determine how "typical" a function's shape is. Like its multivariate counterparts, ACH Depth is highly robust, maintaining its performance even with 30% contamination in the reference set.

Beyond Euclidean Space

The logic of data depth isn't restricted to flat, Euclidean space. In Connecting pairwise spheres by depth: DCOPS, Fabrice Gamboa, Ricardo Fraiman, and Leonardo Moreno extend these concepts to Riemannian manifolds.

By using geodesic distances (the shortest path on a curved surface) to define "spherical" regions of data, they allow researchers to find the center of data clouds on complex structures like spheres, tori, or spaces of positive definite matrices. This ensures that even in non-linear geometries, we can identify outliers and typical patterns without relying on distorted Euclidean approximations.

Summary of the Depth Framework

The shift from parametric modeling to data depth represents a move toward geometric statistics. Whether implemented in the ddalpha toolkit for classification or used for unsupervised anomaly detection, the core mechanism remains the same:

  1. Map the Geometry: Determine the relative positioning of all observations.
  2. Assign Depth: Rank observations from the center outward.
  3. Leverage Stability: Use these rankings to build classifiers or detection thresholds that ignore peripheral noise.

This framework bridges the gap between classical statistical rigor and the practical need for models that don't break when faced with the messy, contaminated reality of modern data.

Papers in this version

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_robust_statistical_learning_via_data_dep_ac21ea9b,
  title = {Robust Statistical Learning via Data Depth},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-statistical-learning-via-data-depth/rdr_ac21ea9b2f904e3986a68d9c0bb62a86},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_robust_statistical_learning_via_data_dep_ac21ea9b_v1,
  title = {Robust Statistical Learning via Data Depth},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-statistical-learning-via-data-depth/rdr_ac21ea9b2f904e3986a68d9c0bb62a86/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}