GeoLatent

Geometry-aware, model-intelligent 3-D visualisations for machine learning workflows.

---

GeoLatent renders the true 3-D decision boundary of any classifier and the geometric structure of any embedding space — in a single function call.

from geolatent import visualize_decision_geometry
from sklearn.svm import SVC

model = SVC(kernel="rbf", probability=True).fit(X, y)
fig = visualize_decision_geometry(model, X, y, projection_method="sensitivity")
fig.show()

Why GeoLatent?

	Standard wrappers	GeoLatent
Projection	Fixed 2-D PCA	PCA · t-SNE · UMAP · Sensitivity
Decision surfaces	Axis-aligned slices	True 3-D isosurfaces via inverse-transform
Confidence regions	None	Nested probability shells + Mahalanobis ellipsoids
Model interface	sklearn only	Any predict / predict_proba callable
Projection axes	Data-driven	Model-driven via finite-difference Jacobians

The sensitivity projection is GeoLatent’s key differentiator: it computes finite-difference Jacobians of any model to find the three directions in feature space where the decision function changes fastest — axes that are task-relevant regardless of input dimensionality.

---

Install

pip install geolatent

With UMAP support:

pip install "geolatent[umap]"

---

Getting Started

• Installation
• Quick Start

User Guide

• Decision Geometry
  • The problem with 2-D slices
  • How GeoLatent does it
  • Projection methods
  • Confidence shells
  • Structural overlays
  • Example
• Latent Space Analysis
  • What it renders
  • Projection methods
  • Example — sentence embeddings
  • Ellipsoid confidence
• Sensitivity Projection
  • Motivation
  • Algorithm
  • Properties
  • Usage
  • Tuning
• Configuration
  • Structure
  • Fluent helpers
  • Manual configuration
  • Custom colour palette
  • ProjectionConfig reference

Tutorials

• Tutorials

Reference

• API Reference
• Changelog

Development

• Contributing to GeoLatent