behavioral analysis

**Behavioral Analysis** of ML models is the **study of model behavior across different input regions, subgroups, and conditions** — going beyond aggregate metrics to understand how the model behaves for different types of inputs, revealing biases, inconsistencies, and failure patterns. **Behavioral Analysis Methods** - **Subgroup Analysis**: Evaluate performance on meaningful subgroups (by tool, product, process window region). - **Error Analysis**: Categorize model errors by type and frequency — identify systematic failure patterns. - **Decision Boundary Exploration**: Probe the model near decision boundaries to understand classification transitions. - **Counterfactual Analysis**: Study how predictions change as individual features are varied. **Why It Matters** - **Failure Patterns**: Aggregate accuracy hides systematic failures on specific subgroups or input types. - **Bias Detection**: Reveals if the model performs differently on different tools, products, or process conditions. - **Process Insight**: Error patterns often reveal insights about the underlying process physics. **Behavioral Analysis** is **understanding the model's personality** — comprehensively studying how it behaves across different situations, inputs, and conditions.