Interpretable Statistical Learning for Asymmetric Decision Costs

Cost-Sensitive Interpretable Modeling

I study models that preserve the explanatory value of statistical coefficients while optimizing directly for asymmetric error costs.

Question

How can we keep model decisions legible when false positives and false negatives have different consequences?

Method

Cost-sensitive logistic regression, structured penalties, and interpretable ensemble construction under non-convex objectives.

Application

Credit scoring, fraud detection, churn prediction, and related business decision problems with unequal misclassification costs.

Representative paper

Ensemble Decision Models and Optimization

I develop ensemble methods that improve decision quality without collapsing into opaque black-box systems.

Question

How can complementary base learners be trained jointly so that the ensemble reduces expected decision cost while remaining understandable?

Method

Diversity-aware objective design and dedicated optimization algorithms, including partial conservative convex separable quadratic approximation (PCCSQA).

Application

Scenarios where robustness and improved cost performance are needed but stakeholders still require a coherent explanation of model behavior.

Published output

Sparse Prescriptive Learning

I investigate prescriptive learning methods that rely on a sparse, representative subset of historical cases instead of the full data archive.

Question

Can prescriptive analytics become both more efficient and more cognitively aligned by selecting a small set of representative prototypes?

Method

Sparse regularization for prototype selection inside prescriptive learning pipelines.

Application

Business decisions where recommendations should be computationally efficient and easier to justify by linking them to representative past examples.