liveBook · Manning

Fairness evaluation

As explained in AI Model Evaluation

A fairness evaluation assesses whether a model's outcomes or error rates differ unfairly across protected or otherwise relevant groups.

Fairness evaluation checks whether model outcomes or error rates differ unfairly across protected or otherwise relevant groups. In practice, it acts as a diagnostic pre-flight check before an A/B test, helping teams catch bias before deployment. section 2.1.2

A practical evaluation starts by defining the outcome, comparing groups, and investigating differences:

  • Define a positive outcome, such as receiving at least one blockbuster or long-tail movie recommendation in the top 10. section 3.2.2
  • Compare outcome rates across groups using demographic parity, then inspect disparity ratios or absolute differences against a reference group. section 3.2.2
  • Use equality of opportunity to compare true-positive rates among qualified users across groups. section 2.1.2 section 3.2.2
  • Treat metric differences as investigation signals rather than automatic proof of unfairness; check user preferences, behavior, product goals, and whether disparities persist after controls. section 3.2.2
Table 2.2 Common Types of Diagnostic Evaluations
Type of Diagnostic Evaluation
Purpose
Example
Error Analysis
Identify patterns in errors (e.g., false positives, false negatives) and their causes.
Error Categorization, Confusion Matrices, Drill-Down by Features
Bias and Fairness Evaluations
Ensure the model provides equitable outcomes and avoids bias across groups.
Group Fairness Metrics, Disparate Impact Testing, Prediction Bias Analysis
Generalization Diagnostics
Understand generalization issues like overfitting, feature contributions, or data shifts.
Bias-Variance Analysis, Feature Importance, Data Drift Analysis
Coverage and Diversity Evaluations
Assess how well the model explores or utilizes the full range of data or options.
Notebooks, access to data layer and visualization tooling.  

Use the diagnostic-evaluation table to distinguish fairness checks from error analysis, generalization diagnostics, and coverage evaluations. section 2.1.2

Movie recommendations and unequal exposure

Suppose a recommender gives different age groups different rates of blockbuster recommendations. A disparity ratio near 1.0 indicates exposure similar to the reference group; values much higher or lower indicate a difference requiring investigation. The same workflow can examine long-tail content or specific categories. Comparable failures can appear in product experiences: Twitter’s image-cropping algorithm was investigated after users reported unequal treatment in preview thumbnails. section 3.2.2

How AI Model Evaluation covers it

How deep it goes
Defines demographic parity and equality of opportunity, shows disparity calculations, and stops short of prescribing universal fairness thresholds.
Examples
Uses blockbuster and long-tail movie recommendations, group disparity ratios, and unequal image-cropping exposure as examples.
What it assumes
Requires defining groups and positive outcomes, while teaching demographic parity and equality of opportunity as starting points.
What you can do after
Diagnose potential group disparities and decide which model behaviors require further investigation before launch.
How it connects
Connects demographic parity and equality of opportunity with error analysis, bias, user preferences, and A/B testing.

Also explained in

Compare how 2 books explain fairness evaluation →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Model Evaluation. The section links point to the parts of the book it draws on.