Learning Credible Models.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Biostatistics

Background:

• Model interpretability is crucial, but models may lack credibility if their reasoning contradicts established knowledge.
• Interpretable models are not always credible, posing challenges in high-stakes applications.
• Defining and achieving model credibility alongside accuracy is an ongoing research problem.

Purpose of the Study:

• To formally define and address model credibility in the linear setting.
• To develop techniques for learning accurate and credible models.
• To introduce a novel regularization penalty that incorporates expert knowledge.

Main Methods:

• Proposed the Expert Yielded Estimates (EYE) regularization penalty.
• Incorporated expert knowledge on covariate-outcome relationships into the penalty.
• Conducted theoretical and empirical comparisons with existing regularization techniques.
• Evaluated methods on synthetic and real-world datasets, including patient risk stratification.

Main Results:

• Models trained with the EYE penalty demonstrated significantly higher credibility compared to other methods.
• The EYE penalty effectively integrates expert knowledge into the model learning process.
• Applied to patient risk stratification, EYE resulted in models with high overlap between top features and known clinical risk factors.
• Achieved good predictive performance concurrently with enhanced credibility.

Conclusions:

• The EYE penalty offers a viable method for enhancing model credibility without sacrificing predictive accuracy.
• This approach is particularly valuable in domains requiring trust and alignment with domain expertise, such as healthcare.
• The findings support the use of expert knowledge integration for building more trustworthy AI systems.