Evidence Geometry is an experimental framework for interpretable risk modeling in classification problems.
Instead of producing a single probability score, it decomposes model predictions into structured evidence signals that reveal how risk emerges in the data.
The framework transforms heterogeneous features into a unified log-likelihood ratio evidence space, allowing risk to be analyzed geometrically.
https://github.com/TheFifthPostulate/evidence-geometry/tree/main
BCW dataset : https://thefifthpostulate.github.io/evidence-geometry/bcw_analysis.html
UCI Heart Disease Cleveland : https://thefifthpostulate.github.io/evidence-geometry/heartdisease_analysis.html
Each feature is converted into log-likelihood ratio evidence
log p(x | positive class) − log p(x | negative class)
This transforms heterogeneous inputs into a unified evidence space.
In this space:
• Evidence accumulates additively
• Population geometry becomes analyzable
• Case-level risk can be decomposed into interpretable components
This interpretation connects the framework to classical likelihood ratio testing and Bayesian evidence accumulation.
Raw Data
↓
Per-Feature Likelihood Models
↓
Log-Likelihood Ratio Evidence
↓
Evidence Space
↓
Geometric Risk Signals
For each case, the model computes:
Difference in Mahalanobis distance relative to learned class manifolds.
→ Measures which class explains the case better in a covariance-aware way.
Projection onto the learned mean deviation separation direction.
→ Captures accumulation of marginal evidence deviations toward the positive class.
Sum of squared projections onto dominant positive-class eigenmodes.
→ Measures activation of characteristic positive-class feature bundles.
Together, these signals capture complementary modes of risk instead of collapsing everything into a single probability score.
A case might produce:
d_dist = -0.3
proj = +0.7
E_pos = moderate
Interpretation:
- globally closer to the benign population
- but drifting toward pathological structure
- with partial activation of disease-related feature bundles
Such cases are natural candidates for review triage rather than automated classification.
- Wisconsin Breast Cancer (BCW)
- UCI Heart Disease (Cleveland)
Across both datasets, the framework:
- Supports interpretable triage
- Reduces false negatives relative to baseline discriminative models
- Maintains automation rate via multi threshold-based review policies
- Fully executable R notebooks
- End-to-end train/validation/test workflow
- Feature-level decomposition utilities
- Principal Component analysis tools
- Modular likelihood specification framework
Core Dependencies: R (4.1 or greater), dplyr, tibble, MASS
Required Packages for Notebooks: dplyr, ggplot2, ranger, caret, patchwork, tibble
Purely discriminative classifiers produce probability scores without exposing internal structure.
Evidence Geometry:
- Preserves per-feature likelihood information
- Separates marginal deviation signals from covariance structure
- Enables threshold-based triage policies grounded in reasoning
- Provides feature bundle interpretation via eigenmodes
This makes it particularly suitable for risk-sensitive domains where interpretability and auditability matter.
Prototype v0.1
Actively expanding to larger clinical datasets (e.g., MIMIC/eICU subsets)