Presentation Profile
Diviner: A Semi-AutoML Approach to Collaborative Model Building
Currently Scheduled: 10/14/2025 - 12:00 PM - 12:45 PM
Room: Tulip
Main Author
Nathanial Watson - Eigenvector Research, Inc.
Abstract:
Diviner is a semi-autoML approach that addresses traditional AutoML's black-box limitations by involving analysts in the model-building process. Instead of producing a single optimal model, the method creates a ranked family of models based on cross-validation performance, overfitting, and prediction error. The process includes user-assisted outlier assessment, variable selection, preprocessing exploration, and linear model calibration. Users can then select models for further refinement. Models can be linear or non-linear. The final output can be a single model, top-ranked models, or an ensemble. This collaborative approach bridges full automation and user customization, delivering improved transparency, interpretability, and model diversity while maintaining predictive accuracy.
