Algorithmic Explanations (WIP)

This project generates explanations of the trained machine learning model. The following steps are included:

• cleaning of data and encoding of the features
• training and tuning a machine learning model (XGBoost)
• generating explanations from four machine learning explainability algorithms with a unified API
• generating different textual and visual explanations for the trained model

Getting Started

Download the github repository:

git clone https://github.com/MauroLuzzatto/algorithmic-explanations

create a local environement and install the python requirements:

pip install -r requirements.txt

Machine Learning explanability methods

Method	Implemenation	Type	Scope	Option
Permutation Feature Importance	PermutationExplanation.py	Non-contrastive	global	sparse and dense
Shapley Values	ShapleyExplanation.py	Non-contrastive	local	sparse and dense
Global Surrogate Model	SurrogateModelExplanation.py	Contrastive	global	sparse and dense
Counterfactual Example	CounterfactualExplanation.py	Contrastive	local	sparse and dense

Explanations

Permutation Feature Importance

Method:

To help you understand this decision, here are the six features which were most important for how the mechanism made its decision in your specific case:

Features:

In your case, the six features which contributed most to the mechanism’s decision were the features 'Number of extracurricular activities' with an average contribution of 0.41, 'Essay score' with an average contribution of 0.17, 'GPA' with an average contribution of 0.08, 'SAT' with an average contribution of 0.05, 'State of residence' with an average contribution of 0.04 and 'Number of activities with leadership experience' with an average contribution of 0.03.

Shapley Values

Method:

To help you understand this decision, here are the six features which were most important for how the mechanism made its decision in your specific case:

Features:

In your case, the six features which contributed most to the mechanism’s decision were the features 'Number of extracurricular activities' with an average contribution of 0.53, 'Essay score' with an average contribution of -0.09, 'GPA' with an average contribution of 0.07, 'State of residence' with an average contribution of 0.06, 'SAT' with an average contribution of -0.05 and 'College rank - unknown' with an average contribution of -0.04.

Counterfactual Example

Method:

To help you understand this decision, here is an example of another, similar applicant where the mechanism would have decided differently:

Features:

In your case, the mechanism would have awarded you the scholarship, if your 'Grade' was High School Senior (12), if your 'Favorite subjects' was not 'Physical Education (P.E.)', if your 'College rank' was not 'unknown', if your 'Favorite subjects' was Chemistry, if your 'Favorite subjects' was Literature and if your 'Favorite subjects' was History.

Global Surrogate Model (Decision Tree)

Method:

To help you understand this decision, here is a decision tree showing you how the mechanism made its decision:

Features:

Applicants received an average score of 2.36 if the value of 'Number of extracurricular activities' is smaller or equal to 3.50 and > 'Number of extracurricular activities' is smaller or equal to 1.50. Applicants received an average score of 2.92 if the value of 'Number of extracurricular activities' is smaller or equal to 3.50 and > 'Number of extracurricular activities' is larger than 1.50. Applicants received an average score of 3.56 if the value of 'Number of extracurricular activities' is larger than 3.50 and 'SAT' is > smaller or equal to 1210.00. Applicants received an average score of 4.29 if the value of 'Number of extracurricular activities' is larger than 3.50 and 'SAT' is > larger than 1210.00.

Authors

• Mauro Luzzatto - Maurol