A supervised clustering MCMC methodology for large categorical feature spaces

Area of Science:

• Statistics
• Machine Learning
• Health Informatics

Background:

• Dimensionality reduction is crucial in machine learning, often addressed via unsupervised methods.
• Large categorical feature spaces pose challenges for predictive modeling.
• Existing risk adjustment models in health insurance can be improved.

Purpose of the Study:

• To introduce a novel supervised clustering methodology for optimizing large categorical feature spaces.
• To enhance the accuracy of supervised learning algorithms by minimizing test error.
• To improve risk adjustment in health insurance markets through better health expenditure prediction.

Main Methods:

• Developed a supervised clustering methodology utilizing a Metropolis Hastings algorithm.
• Optimized the partition structure of large categorical feature spaces.
• Applied the methodology to cluster International Classification of Diseases, 10th Revision (ICD-10) codes.

Main Results:

• The supervised clustering approach demonstrated superior performance compared to common alternatives.
• Clustering diagnostic codes into risk groups improved health expenditure prediction.
• The methodology was validated on a large dataset from the Colombian Healthcare System.

Conclusions:

• The proposed supervised clustering method effectively reduces dimensionality and minimizes test error.
• This approach offers a significant improvement for risk adjustment in health insurance.
• The methodology has broad applicability to supervised learning problems with large categorical features.