Zoish: A Novel Feature Selection Approach Leveraging Shapley Additive Values for Machine Learning Applications in...

Area of Science:

• Healthcare Analytics
• Machine Learning
• Cooperative Game Theory

Background:

• Effective feature selection is critical for robust predictive models in healthcare analytics.
• Challenges include small sample sizes and potential data biases.
• Existing tools often lack transparency and automation in feature selection.

Purpose of the Study:

• To introduce Zoish, a tool for transparent and automated feature selection using Shapley values.
• To demonstrate Zoish's versatility across various machine learning libraries and dataset sizes.
• To enhance interpretability and user control in healthcare predictive modeling.

Main Methods:

• Employs Shapley additive values derived from cooperative game theory for feature selection.
• Features a dual algorithmic approach for efficient Shapley value calculation on large and small datasets.
• Integrates seamlessly with scikit-learn, XGBoost, CatBoost, and imbalanced-learn.

Main Results:

• Zoish provides transparent and automated feature selection, improving model robustness.
• Demonstrated adaptability and efficiency in case studies for breast cancer and Parkinson's disease (Montreal Cognitive Assessment - MoCA) prediction.
• Evaluated on 300 synthetic datasets, showing superior performance compared to counterparts.

Conclusions:

• Zoish offers a unique, streamlined process combining local and global feature selection.
• Its interpretability features and customizable settings optimize predictive objectives.
• Zoish is highly suitable for diverse healthcare analytics tasks, enhancing predictive model development.