Slice-Based Evaluation of ML Models

Description

Evaluate a model separately on meaningful subsets of data to catch hidden subgroup failures, weak critical slices and Simpson's paradox reversals.

Overall metrics can hide a model that performs badly on a minority or more critical subgroup, and aggregation can reverse comparisons. Huyen's example: model A scores 98% and 80% on two groups (96.2% overall) and model B 95% on both, so A looks better overall but fails the second group. Slices come from three sources: domain heuristics (mobile versus desktop, browser, location, new versus returning users), error analysis by reading misclassified examples for patterns, and automated slice finders that generate candidates by beam search, clustering or decision trees, then prune and rank them. Each critical slice needs enough correctly labelled evaluation data. Slicing can also expose problems that are not about the model at all, such as poor results on mobile traced to a half-hidden button on small screens.

Attributes

Loss type
Siloed
Effort
Days

Steps

  1. Identify candidate slices from domain knowledge, including user groups, devices, locales and high-stakes cases.
  2. Read misclassified examples to spot patterns that suggest further slices.
  3. Optionally run an automated slice finder to generate, prune and rank slices by error and size.
  4. Collect enough correctly labelled evaluation data for each critical slice.
  5. Report performance per slice alongside the overall figure, and check for Simpson's paradox reversals between models.
  6. Choose or improve models using sliced results, and investigate non-model causes such as interface problems.

Sources

  • Designing Machine Learning Systems (Huyen, 2022)

Also known as

  • Slice-based evaluation
  • Subgroup evaluation
  • Sliced metrics
  • Slice Finder

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning