Model behavior is the observable outputs and actions produced in response to inputs and conditions. Evaluating behavior means looking beyond overall scores to understand what the model does for particular users, feature values, and edge cases. section 2.1.2
Diagnostic evaluations investigate why behavior occurs, expose weaknesses and bias, and verify whether the model fulfills its objective before release or A/B testing. Strong aggregate performance can still hide catastrophic failures for specific groups or unusual situations. section 2.1.2 section 3.1
Useful diagnostic views of behavior include:
- Error analysis finds patterns in misclassifications, false positives, false negatives, and predictions that violate expectations. section 2.1.2
- Bias and fairness evaluations examine whether outcomes systematically advantage or disadvantage demographic groups, using measures such as demographic parity or equality of opportunity. section 2.1.2
- Generalization diagnostics investigate overfitting, feature contributions, and data shifts through methods such as feature-importance and data-drift analysis. section 2.1.2
- Coverage and diversity evaluations assess whether the model explores or uses the available range of data or options. section 2.1.2
A hidden group disparity
A model may perform well overall while predicting lower-quality candidates for a particular demographic. Group-level diagnostics can surface the disparity, while error analysis helps identify the associated prediction failures before they harm product users. section 2.1.2 section 3.1
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.
|
Compare diagnostic evaluation types with their purposes and practical examples, including confusion matrices, disparate-impact testing, feature-importance analysis, and data-drift analysis. section 2.1.2





