Machine Learning Interpretability

1. Explainer Selection: Choose the appropriate SHAP explainer for the model architecture (TreeExplainer for tree-based models, DeepExplainer for neural networks, KernelExplainer for model-agnostic/black-box). 2. Background Data: For explainers requiring background data, select a representative sample to establish the 'base value' (expected model output) while minimizing compute cost. 3. Local Explanation: Use waterfall or force plots to explain individual predictions. Ensure the sum of SHAP values plus the base value equals the model's actual output. 4. Global Explanation: Use beeswarm or summary plots to evaluate overall feature importance and directionality across the dataset. 5. Interaction Analysis: Use dependence plots to observe how a single feature affects the output and how it interacts with other features. 6. Common Mistakes to Avoid: Never interpret SHAP values as causal effects (they only explain the model's learned associations, not reality); do not ignore feature correlation, which can distort attribution by evaluating unrealistic data combinations; and do not confuse the base value with a zero-baseline.

Grounding (do this at runtime)

Do NOT answer precise facts (codes, identifiers, sections, current values) from memory — run the bundled script, and base the answer only on what it returns.

NOTICE

Based on the SHAP framework introduced by Lundberg & Lee (2017) and best practices from Christoph Molnar's 'Interpretable Machine Learning'.