Explanations Based on Item Response Theory (eXirt): A Model-Specific Method to Explain Tree-Ensemble Model in Trust Perspective

eXirt offers a specialized approach to explaining tree-ensemble models used for binary classification on tabular data, addressing a gap where existing explanation methods (SHAP, DALEX, etc.) lack consensus on which is most appropriate for this common type of model.

The method combines two types of explanations for better trust assessment: global feature rankings that show which inputs matter most overall, plus local "explanation-by-example" approaches that help users understand why specific predictions were made, providing a more complete picture of model behavior.

Extensive empirical validation across 41 datasets and 4 major algorithms demonstrates that eXirt's feature importance rankings are comparable to established XAI methods, suggesting it's a reliable alternative rather than an outlier approach.

Item Response Theory (IRT) provides a novel theoretical foundation for model explanations, borrowing from educational testing methodology to quantify how model features respond to different input conditions, offering a different perspective than existing popularity-driven XAI methods.

The focus on tree-ensemble models addresses practical needs since these algorithms (Random Forest, Gradient Boosting, CatBoost, LightGBM) are widely used in real-world applications, making specialized explanation methods directly applicable to systems already deployed in industry.