This repository demonstrates a likelihood-based approach for transforming heterogeneous features into a common, comparable representation using log-likelihood ratios (LLR).
The goal is to standardize feature values relative to outcome-specific distributions and enable interpretable analysis and modeling across datasets.
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
For each feature, values are transformed into a log-likelihood ratio:
log p(x | positive) − log p(x | negative)
This converts raw inputs into a common evidence scale, allowing:
- comparison across heterogeneous variables
- additive aggregation of feature contributions
- analysis of samples relative to class-specific distributions
The transformed representation can be summarized using complementary signals:
Difference in Mahalanobis distance to class-specific distributions
→ captures relative proximity under covariance structure
Projection onto the primary separation direction
→ captures directional accumulation of feature-level deviations
The framework is demonstrated on:
- Breast Cancer Wisconsin (BCW)
- UCI Heart Disease (Cleveland)
These examples illustrate:
- transformation of raw features into a comparable representation
- extraction of interpretable summary signals
- applicability across labeled datasets
- R notebooks with end-to-end workflows
- feature transformation utilities (LLR-based)
- signal extraction and visualization tools
install.packages(c( "dplyr", "ggplot2", "ranger", "caret", "MASS", "patchwork" ))