Evaluation methods

• Perturbation tests
  • make small changes to your test splits to see how these changes affect your model’s performance
• In-variance tests
  • certain changes to the inputs shouldn’t lead to changes in the output.
• Directional expectation tests
  • e.g.if x increases, y is expected to decrease; if not, you should check before deploying it
• Model calibration
  • a common method is Platt scaling
• Confidence measurement
  • think about the usefulness threshold for each individual prediction
• Slice-based evaluation
  • consider Simpson’s paradox
  • issues
    • model performs differently on different slices of data when the model should perform the same
    • model performs the same on different slices of data when the model should perform differently
  • thus, you should discover critical slices in your data using
    • heuristics-based approach: slice the data using domain knowledge
    • Error analysis: manually go through misclassified examples and find patterns among them
    • slice finder

Robustness across resampling

• Use Resampling-based Model Stability Checks to test whether model performance is robust across different train/test splits, folds, bootstrap samples, or subsamples.
• This helps distinguish a reliable model from one that performed well because of one lucky split.