Can a Transparent Machine Learning Algorithm Predict Better than Its Black Box Counterparts? A Benchmarking Study...

Area of Science:

• Computer Science
• Statistics
• Artificial Intelligence

Background:

• Black box machine learning (ML) models often lack transparency, hindering human understanding.
• Existing ML algorithms struggle to balance model interpretability with flexibility for nonlinearity and interactions.

Purpose of the Study:

• To develop a novel, human-centered ML algorithm that produces transparent models.
• To evaluate the performance of this new algorithm against popular black box methods.

Main Methods:

• Developed a novel algorithm based on ranked sparsity, prioritizing simpler terms over complex interactions.
• Implemented the algorithm in the open-source R package, sparseR.
• Benchmarked the algorithm against other ML methods on simulated and real-world datasets from the Penn Machine Learning Benchmarks database.

Main Results:

• The human-centered algorithm achieved competitive predictive accuracy, rivaling black box approaches like neural networks, random forests, and support vector machines.
• Interpretable approaches performed optimally or within 5% of the optimal method on most real-world datasets.
• Performance was comparable to black box methods, with some cases of underperformance noted for interpretable methods.

Conclusions:

• The novel ranked sparsity algorithm offers a viable alternative to black box models, providing both interpretability and competitive accuracy.
• Human-centered transparent algorithms should be considered for predictive modeling applications.
• The sparseR package facilitates the use of interpretable ML methods.