New Algorithm Quantifies Sensitivity of Decision Tree Ensembles
Researchers have developed a novel algorithmic technique to quantify the sensitivity of decision tree ensembles (DTEs), a popular AI model used in safety-critical domains. The method discretizes the input space and enumerates regions susceptible to misclassification from small feature changes. It encodes the problem as an algebraic decision diagram (ADD) and splits it into compositional subproblems, achieving efficient computation within certified error and confidence bounds. The work addresses a key verification challenge for DTEs, which are widely deployed in high-stakes applications.
Key facts
- Decision tree ensembles are used in multiple safety critical domains.
- Sensitivity asks whether a small change in subset of features can lead to misclassification.
- The approach discretizes the input space and enumerates sensitive regions.
- The algorithm uses an algebraic decision diagram (ADD) encoding.
- The computation is compositional and splits into subproblems.
- Certified error and confidence bounds are provided.
- The work focuses on a quantitative notion of sensitivity tailored to DTEs.
- The technique is novel and efficient.
Entities
Institutions
- arXiv