A Bayesian Approach to Learning Scoring Systems

Area of Science:

• Statistical modeling
• Machine learning
• Bayesian inference

Background:

• Traditional scoring systems often rely on manual heuristics or suboptimal coefficient scaling.
• Existing methods for constructing scoring systems can lead to reduced interpretability and performance.

Purpose of the Study:

• To develop a data-driven Bayesian method for constructing interpretable scoring systems.
• To automate the creation of scoring systems by learning from data with specified priors.

Main Methods:

• A Bayesian approach is proposed, utilizing a Metropolis-Hastings sampler.
• The method learns scoring systems by specifying priors on coefficient characteristics.
• Coefficients are learned from data and guided towards a 'natural scale'.

Main Results:

• The proposed method yields highly interpretable models.
• The learned scoring systems demonstrate competitive generalization performance.
• This approach overcomes limitations of manual heuristic-based scoring system construction.

Conclusions:

• The Bayesian method offers an effective way to build interpretable and high-performing scoring systems.
• This data-driven approach enhances the reliability and applicability of scoring systems in various fields.
• Future work could explore extensions to more complex model structures.