A note on two novel easy-to-interpret feature effect measures for partial dependence plots in a classification...

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Parametric models like logistic regression are traditional for classification tasks.
• Black box supervised learning models (BBSLMs) often exceed parametric models in prediction accuracy.
• BBSLMs lack easily interpretable feature effect measures comparable to logistic regression's odds ratio (OR).

Purpose of the Study:

• To derive novel feature effect measures for binary classification using BBSLMs.
• To introduce the relative risk of marginal effects (RRME) and odds ratio of marginal effects (ORME).
• To assess the performance and interpretability of these new measures.

Main Methods:

• Derivation of RRME and ORME based on partial dependence plots for BBSLMs.
• Application to a dataset of patients admitted to hospital with myocardial infarction.
• Comparison of BBSLMs with RRME/ORME against traditional logistic regression models.

Main Results:

• BBSLMs demonstrated superior predictive ability compared to logistic regression.
• The derived RRME and ORME for anterior infarct (a key risk factor) were 1.8, comparable to the logistic regression OR of 1.9.
• RRME and ORME proved more robust, handling observations with missing feature values effectively.

Conclusions:

• The novel RRME and ORME measures provide interpretable feature effects for BBSLMs in binary classification.
• These measures offer comparable interpretability to traditional OR while leveraging the superior predictive power of BBSLMs.
• The robustness of RRME and ORME to missing data enhances their practical applicability in statistical modeling.