Model LineUpper: Supporting Interactive Model Comparison at Multiple Levels for AutoML
summary
Model LineUpper is a prototype that combines multiple XAI techniques and visualizations — a metrics table for elimination, a probability scatter matrix for pairwise model-instance comparisons, and global plus local feature-importance views showing mean local feature contribution. It runs on IBM AutoML, produces four variants per algorithm by iterating hyperparameter tuning and feature engineering, and is evaluated with 14 data scientists. Users find the comparisons useful but report confusion over transformed feature names and want more raw-data statistics; recommendations push for faster interactive comparisons and grouped algorithm views.
AI-written summary · about
chapters (9)
- 0:00Start
- 0:23The current state of AutoML
- 1:07The Model LineUpper
- 1:33Metrics & Algorithm Variations
- 2:12Probability Scatterplot Matrix
- 3:04Local Instances & Feature Contribution
- 3:24User Study Results
- 3:53Design Recommendations
- 4:35Conclusion & Future Work
