Random forest versus logistic regression: a large-scale benchmark experiment.

Area of Science:

• Machine Learning
• Computational Biology
• Data Science

Background:

• The Random Forest (RF) algorithm has become a popular classification tool since 2001.
• It is increasingly used in scientific fields, competing with logistic regression (LR).

Purpose of the Study:

• To benchmark the prediction performance of the original RF algorithm against LR.
• To compare these algorithms using a large-scale experiment with 243 real datasets for binary classification.

Main Methods:

• A large-scale benchmarking experiment was conducted using 243 real-world datasets.
• The study compared the original Random Forest (RF) algorithm with default parameters against logistic regression (LR).
• The experimental design was inspired by clinical trial methodology to minimize bias.

Main Results:

• Random Forest demonstrated superior accuracy in approximately 69% of the datasets compared to logistic regression.
• Statistical analysis showed significantly better performance for RF across accuracy, Area Under the Curve (AUC), and Brier score.
• Dataset selection criteria notably influenced the observed results, underscoring the importance of transparent methodology.

Conclusions:

• The original Random Forest algorithm with default parameters generally outperforms logistic regression for binary classification.
• Future research should involve neutral, large-scale studies to evaluate various RF implementations and parameters.
• Clear reporting of dataset selection criteria is crucial for reproducible and reliable algorithm benchmarking.