Integer Programming Models for Interpretable Binary Classification
Abstract
Discrete optimization techniques have been successfully applied to many industrial applications such as production planning, scheduling and logisics. However, these techniques, and especially integer linear programming (IP), is not yet widely used in AI applications even for those that are combinatorial in nature (eg., classification or clustering). An important reason behind this is the presence of large data sets in AI applications which usually lead to large-scale IPs that are hard to solve. Recently, applying IP techniques to some AI problems has become an active area of research. Some of these problems are: Decision Trees [10, 40, 72, 63, 51, 2], Decision Rule Sets [53, 59, 21], and, Adversarial Neural Networks with Relu Gates [35, 4].With the use of machine learning (ML) techniques to automate socially and legally sensitive decision making tasks such as lending, hiring, and college admissions, the need for interpretable ML models is increasing. In these applications ML models typically complement human decision makers and consequently transparency of the model is crucial in order for the decision maker to understand, critique and trust this model. Interpretable ML is an area well suited for discrete optimization as interpretable models are often well-modeled by low-complexity discrete objects. In these applications, low model complexity is considered as a proxy for interpretability (the degree to which a human can understand the cause of a decision). As one would expect, there is usually a trade-off between interpretability/complexity and predictive accuracy.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2021
- Source ID
- N000142112575
Entities
People
- Oktay Gunluk
Organizations
- Cornell University
- Office of Naval Research
- United States Navy