Beyond XGBoost and SHAP: Unveiling true feature importance

Area of Science:

• Machine Learning
• Data Science
• Statistical Modeling

Background:

• XGBoost and SHAP values are powerful tools for machine learning analysis.
• However, they can introduce analytical pitfalls, such as inflated feature importance and interpretation biases.
• Lack of ground truth complicates rigorous model evaluation.

Purpose of the Study:

• To outline key machine learning principles for researchers.
• To identify and address potential biases in XGBoost and SHAP value interpretation.
• To advocate for rigorous statistical methods in machine learning research.

Main Methods:

• Focus on XGBoost's incremental decision tree building process.
• Analysis of SHAP values' dependence on model structure and feature interactions.
• Discussion on the role and limitations of ground truth values in model evaluation.

Main Results:

• XGBoost may inflate feature importance due to its focus on misclassified examples.
• SHAP values can be biased by model structure and feature interactions.
• Ground truth values are essential for accuracy but do not guarantee real feature-target associations.

Conclusions:

• Researchers must be aware of biases in feature importance derived from methods like XGBoost and SHAP.
• Rigorous statistical methods are necessary for reliable model evaluation and interpretation.
• Understanding these limitations enhances the credibility of machine learning research findings.