Semi-Automated Machine Learning for Calibration Model Development

This talk explores Diviner, a semi-automated machine learning approach for calibration model development presented by Manuel A. Palacios and Barry M. Wise from Eigenvector Research. Learn how Diviner differs from traditional AutoML by actively involving analysts in the model-building process instead of producing a single black-box solution. Discover how this approach creates and ranks a family of models based on cross-validation performance, overfitting assessment, and prediction error on validation sets. The presentation details Diviner's workflow, which includes user-assisted outlier assessment through Principal Components Analysis and Robust Partial Least Squares, variable selection, exploration of preprocessing algorithms, and calibration of linear models with hyperparameter tuning. See how this collaborative approach bridges the gap between full automation and user-led customization, resulting in more transparent and insightful model-building. Through evaluation of multiple datasets, the speakers demonstrate Diviner's effectiveness in achieving both predictive accuracy and model diversity, moving machine learning calibration away from black-box solutions toward a more interpretable and adaptive methodology.