Skip to content

A model with dichotomized features for explaining binary output

Notifications You must be signed in to change notification settings

lapkin25/XAI-Model

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

XAI-Model

  • Model with dichotomized features: finding the cut-off thresholds and weights
  • Augmentation of model features: adding combinations of features which are binarized
  • Refitting the intercept when weights are fixed
  • Model with piecewise weights
  • Mixed model of a logistic regression and a generalized additive model

Algorithms

Constructing baseline model

The baseline model includes binarized individual features. The weights and cut-off thresholds are found by the following way.

The first step is to construct univariate models for each feature. To this end, we find a weight and a cut-off threshold. The weight is equal to width of border area, where suspected points are situated. At the given weight, we have a border area Π. Let define a property Φ of points at the border area. This property is selected to predict the "1" class at the border area. More precise, the relation TPV/FPV (at the area Π∩Φ) is to be as high as possible.

Improving the model

At the second step, we have the multivariate model, and feature-wise model enhancement can be applied. Let remove an individual feature which is selected randomly and add it again to the model with adjusted weight and cut-off threshold. This procedure is repeated until the stabilization of coefficients.

Augmentation of model features

Border area method is proposed here to create new binary features which are determined by a few features.

Explainability

The border area method can be considered as interpretable. When a selected feature is considered, its border area contains the points for which this feature is crucial.

About

A model with dichotomized features for explaining binary output

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages