AIGrammar

AIGrammar

About

AIGrammar is all in one and easy to use package for model diagnostic and vulnerability checks. It enable with a simple line of code to check model metrics and prediction generalizability, feature contribution, and model vulnerability against adversarial attacks.

Data

• Multicollinearity
• Data drift

Model

• Metric metric comparison
  • roc_auc vs average precision
• Optimal threshold vs 50% threshold

Feature importance

• Too high importance
• 0 impact
• Negative influence (FLOFO)
• Causes of overfitting

Adversarial Attack

• Model vulnerability identification based on one feature minimal change for getting opposite outcome.

Usage Python 3.7+ required. Installation: pip install AIGrammar

Example:

aig = AIGrammar(train, test, model, target_name)
aig.measure_all(X0_shap_values, X1_shap_values)

print(aig.diagnosis)
print(aig.warnings)