Adaptable Interpretable Machine Learning - Recursive Bayesian Rule Lists: FY17 Line-Supported Information, Computation, and Exploitation Program
Abstract
Three important factors that influence user trust in automation and machine learning (ML) algorithm are good performance, interpretability-the ability for users to understand how the automation reaches its recommendation-and adaptability-the ability for the automation to learn from new data and user feedback. This report describes work conducted under the Adaptable Interpretable Machine Learning (AIM) program, whose goal is to create ML algorithms that users can understand and that keep learning so users will trust them and actually use them. This report derives Recursive Bayesian Rule Lists (RBRL), a group of supervised-classification algorithms that are both interpretable and adaptable. RBRL is based on Bayesian Rule Lists (BRL), which are interpretable decision-list classifiers that perform competitively with state-of-the art, non-interpretable classifiers on many problems, and on an analogy between classifier adaptation and recursive Bayesian tracking (RBT).
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 12, 2017
- Accession Number
- AD1098167
Entities
People
- J. K. Su
Organizations
- MIT Lincoln Laboratory