Softening the Prediction Surface of Decision Tree Regressors
Abstract
While decision trees function as base-learners for many machine learning methods and exhibit several useful properties, the compromise between prediction accuracy and generalization limits their utility in non-ensemble applications. We propose a new method to improve the predictive performance of pre-trained decision tree regressors. Using the tree structure and metadata, we derive a set of decision threshold-based weights that modify the leaf prediction values. The weighted values are then aggregated into a final softened prediction which more accurately represents the true target distribution. We demonstrate the approach on a variety of benchmark data sets and observe a mean improvement of 11 percent over the baseline decision tree R2 values. We further explore the parameters of the approach and characterize their effects.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 10, 2024
- Accession Number
- AD1229938
Entities
People
- Andrew B. Sabater
- Benjamin A. Michlin
- Jamal T. Rorie
- Joshua A. Duclos
Organizations
- Naval Information Warfare Center Pacific