Multiple Boosting Calibration Trees (MBCT) is a feature-aware binning framework designed to improve the calibration of machine learning classifiers. Traditional classifiers focus on accuracy, but certain applications require calibrated probability estimates. Existing binning methods have limitations, such as only considering original prediction values and being non-individual. MBCT addresses these issues by optimizing the binning scheme using tree structures of features and adopting a linear function in a tree node for individual calibration. MBCT is non-monotonic and improves order accuracy due to its learnable binning scheme. Experiments on various datasets show that MBCT outperforms competing models in calibration error and order accuracy, with a novel multi-view calibration loss as a better metric for modeling calibration error.
Boosting, calibration trees
Tree-based models
Various datasets for calibration
Calibration error, order accuracy
Research environments
No
Yes
Feature-aware binning, individual calibration
No
Standard computing resources
Linux, Windows, macOS
Compatible with tree-based models
N/A
N/A
N/A
Yes
Research community
N/A
Varies based on application
Depends on model complexity
Standard for tree-based models
N/A
N/A
Focus on specific calibration tasks
AI research, medical, meteorological, advertising
Improving model calibration
Researchers
Integrates with tree-based models
Scalable with model size
Community support
N/A
Command-line
No
N/A
Open-source
Yes
Research institutions
N/A
N/A
N/A
Research tool
No
N/A
Open-source
0.00
N/A
Open-source
01/01/1970
01/01/1970
N/A
N/A
Yes