LLpowershap: logistic loss-based automated Shapley values feature selection method.

Area of Science:

• Machine Learning
• Bioinformatics
• Data Science

Background:

• Shapley values are crucial for explaining complex machine learning models, aiding in debugging, fairness analysis, and feature selection.
• Existing methods like powershap utilize predictive Shapley values and p-values for feature selection.
• Shapley values ensure fair distribution of feature contributions, considering non-linearities and interactions.

Purpose of the Study:

• Introduce LLpowershap, a novel feature selection method utilizing loss-based Shapley values.
• Enhance the identification of informative features while minimizing noise.
• Improve the calculation of p-values and power for feature selection and iteration estimation.

Main Methods:

• Employs loss-based Shapley values for feature selection.
• Incorporates enhanced p-value and power calculations.
• Evaluated through simulations and benchmarking on real-world datasets.

Main Results:

• LLpowershap identifies a greater number of informative features and fewer noise features compared to existing methods.
• Demonstrates high or comparable predictive performance against other Shapley-based and filter methods.
• Achieved the top mean ranking among seven tested feature selection methods.

Conclusions:

• LLpowershap is an effective wrapper feature selection method.
• Suitable for feature selection in large-scale biomedical datasets and other complex data environments.
• Offers a robust approach for identifying key features in machine learning models.