A Novel Algorithm to Estimate the Significance Level of a Feature Interaction Using the Extreme Gradient Boosting...

Area of Science:

• Biostatistics
• Machine Learning
• Computational Biology

Background:

• Interaction effects are crucial in cardiac research, but conventional analyses may yield erroneous conclusions if not properly addressed.
• Statistical methods like regression can model interactions, but machine learning approaches often lack p-value assessment for specific feature interactions.

Purpose of the Study:

• To propose a novel machine learning algorithm, XGB-FI (extreme gradient boosting machine for feature interaction), for assessing the p-value of feature interactions.
• To address the limitation of existing machine learning strategies in quantifying the statistical significance of feature interactions.

Main Methods:

• XGB-FI stratifies data into four subgroups based on interactive features.
• It employs cross-validation to train four XGBoost models, preventing overfitting.
• A novel feature interaction ratio (FIR) is computed, and an empirical p-value is derived from its distribution.

Main Results:

• Computer simulations confirmed XGB-FI's valid type I error rate at the 0.05 nominal level.
• XGB-FI consistently exhibited higher statistical power than multiple regression models across all examined scenarios.
• The proposed algorithm effectively identifies significant interaction effects.

Conclusions:

• The novel XGB-FI algorithm outperforms conventional statistical models in detecting feature interactions.
• XGB-FI provides a robust machine learning solution for evaluating the statistical significance of feature interactions in complex datasets.