Divide et Calibra: Multiclass Local Calibration via Vector Quantization
Abstract
Accurate and well-calibrated Machine Learning (ML) models are mandatory in high-stakes settings, yet effective multiclass calibration remains challenging: global approaches assume calibration errors are homogeneous across the latent space, while local methods often rely on representation-space dimensionality reduction, discarding latent directions relevant to miscalibration. To address these issues, we propose a compositional approach to multiclass calibration, where region-specific calibration maps are constructed from shared factors. We instantiate this idea via Vector Quantization (VQ), which induces a structured partition of the representation space, and an indexed parameterization of Dirichlet concentrations that enables parameter sharing across regions. Our approach learns heterogeneous calibration maps that generalize well even to sparse regions of the latent space. Experiments on benchmark datasets show significant improvements in local calibration while maintaining competitive global calibration and predictive performance.
Related
- Hierarchical Utility Calibration for Structured Multiclass Decisions
- Evaluation of TF-IDF and Word2Vec Representations Using Support Vector Machine for Multiclass Product Category Classification of Tokopedia Reviews
- Beyond Expert Rankings: Contrast Calibration for Mixed-Precision MoE Quantization
- SSTQ:Privacy-Preserving Vector Quantization via Subsampled Stochastic TurboQuant
- Prior-Aided Masked Vector Quantization CSI Feedback for FDD Massive MIMO Systems
BibTeX
@misc{barbera2026divide,
title = {{Divide et Calibra: Multiclass Local Calibration via Vector Quantization}},
author = {Cesare Barbera and Lorenzo Perini and Giovanni De Toni and Andrea Passerini and Andrea Pugnana},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2605.21060},
eprint = {2605.21060},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2605.21060},
note = {CC BY-NC-SA 4.0},
}
RIS
TY - UNPB
TI - Divide et Calibra: Multiclass Local Calibration via Vector Quantization
AU - Cesare Barbera
AU - Lorenzo Perini
AU - Giovanni De Toni
AU - Andrea Passerini
AU - Andrea Pugnana
PY - 2026
DA - 2026/10/26
PB - arXiv
DO - 10.48550/arxiv.2605.21060
UR - https://arxiv.org/abs/2605.21060
AB - Accurate and well-calibrated Machine Learning (ML) models are mandatory in high-stakes settings, yet effective multiclass calibration remains challenging: global approaches assume calibration errors are homogeneous across the latent space, while local methods often rely on representation-space dimensionality reduction, discarding latent directions relevant to miscalibration. To address these issues, we propose a compositional approach to multiclass calibration, where region-specific calibration maps are constructed from shared factors. We instantiate this idea via Vector Quantization (VQ), which induces a structured partition of the representation space, and an indexed parameterization of Dirichlet concentrations that enables parameter sharing across regions. Our approach learns heterogeneous calibration maps that generalize well even to sparse regions of the latent space. Experiments on benchmark datasets show significant improvements in local calibration while maintaining competitive global calibration and predictive performance.
ER -