8.3 Modeling Techniques for Interpretability Requirements
Key Takeaways
- Intrinsically interpretable models such as linear and logistic regression expose weights directly, while tree ensembles and DNNs need post-hoc explanation methods.
- Agent Platform feature attribution methods are sampled Shapley for non-differentiable models, integrated gradients for differentiable models, and XRAI for image regions.
- BigQuery ML explains boosted tree and random forest models with Tree SHAP or approximate feature contribution, and DNNs with integrated gradients.
- Feature attributions are measured relative to a baseline, so the choice of baseline changes how attributions should be read.
- Vertex Explainable AI was deprecated on March 16, 2026, and Google points to open-source SHAP and LIME as alternatives.
The exam guide lists modeling techniques given interpretability requirements as part of building models. Interpretability can be a legal requirement (explaining a credit denial), a trust requirement (clinicians must see why), or a debugging aid. This section is about choosing the technique. Implementing and monitoring explanations in production is covered in Section 18.4.
Two Paths to Interpretability
| Path | Examples | Pros | Cons |
|---|---|---|---|
| Intrinsically interpretable ("glass-box") models | Linear or logistic regression, shallow decision trees, rule lists, ARIMA_PLUS decomposition | Explanations are exact and stable. Easy to audit | Can lose accuracy on complex non-linear data |
| Post-hoc explanations of complex models | Shapley-based feature attributions for tree ensembles and DNNs, saliency for images, example-based explanations | Keep high accuracy | Approximations that depend on method and baseline. Harder to validate |
Decision rule: if a regulator or policy requires a precise, stable reason for each decision and the accuracy gap is small, prefer a glass-box model. If complex models are much more accurate and approximate explanations are acceptable, use a complex model plus post-hoc attributions, and validate that the explanations make sense.
Local vs. Global Explanations
- Local explanations show how each feature contributed to one prediction ("income and debt ratio lowered this applicant's score"). They're needed for individual adverse-action reasons.
- Global explanations show a feature's overall influence across the dataset. They're used for model validation and governance reviews.
Explanation Methods on Google Cloud
Agent Platform feature attributions
All three methods are based on Shapley values, which credit each feature for its share of the outcome.
| Method | Best for | Notes |
|---|---|---|
| Sampled Shapley | Non-differentiable models such as tree ensembles, and meta-ensembles of trees and neural networks (AutoML tabular) | Sampling approximation of exact Shapley values. Works with any custom-trained model in any container |
| Integrated gradients | Differentiable models (neural networks), especially with large feature spaces. Low-contrast images such as X-rays | Integrates gradients along a path from a baseline to the input |
| XRAI | Image classification with natural images | Builds on integrated gradients, then oversegments the image and ranks regions by attribution |
Example-based explanations return the most similar training examples, found by nearest-neighbor search on embeddings. They help debug mislabeled or underrepresented data and flag inputs far from anything seen in training. They support TensorFlow models that produce embeddings, not tree models.
BigQuery ML explanations
| Model | Method | Functions |
|---|---|---|
| Linear and logistic regression | Shapley values (weight × standardized feature), plus standard errors and p-values | ML.EXPLAIN_PREDICT, ML.GLOBAL_EXPLAIN, ML.ADVANCED_WEIGHTS |
| Boosted trees, random forest | Tree SHAP (exact for trees), approximate feature contribution, Gini-based importance | ML.EXPLAIN_PREDICT, ML.GLOBAL_EXPLAIN, ML.FEATURE_IMPORTANCE |
| DNN, Wide & Deep | Integrated gradients | ML.EXPLAIN_PREDICT, ML.GLOBAL_EXPLAIN |
| AutoML Tables | Sampled Shapley | ML.GLOBAL_EXPLAIN |
| ARIMA_PLUS | Time series decomposition | ML.EXPLAIN_FORECAST |
Baselines matter
Attributions measure contribution relative to a baseline, such as the median input for tabular data or an all-black image. Change the baseline and the attributions change. Pick a baseline that fits the question, like "compared with a typical applicant."
Lifecycle Note: Explainable AI Deprecation
Vertex Explainable AI was deprecated on March 16, 2026 and is scheduled to shut down on March 16, 2027. No new features are being added, and Google recommends open-source libraries such as SHAP and LIME as alternatives. For long-lived designs:
- Prefer glass-box models where they meet accuracy needs.
- Use BigQuery ML explanation functions for BigQuery ML models.
- Package SHAP (Shapley-based, including TreeSHAP for trees) or LIME (local surrogate models) in training, batch, or serving code for custom models.
Techniques That Improve Interpretability Without Losing Much Accuracy
- Feature selection: fewer, meaningful features make any explanation clearer (Chapter 2).
- Domain-meaningful feature engineering: "debt-to-income ratio" explains better than 40 raw balance columns.
- Constraints: limit tree depth, or use monotonic relationships where the domain requires them ("higher income never lowers the score").
- Two-stage designs: a glass-box model for the regulated decision, and a complex model only for non-regulated ranking.
- Model documentation: record intended use, training data, evaluation slices, and limitations.
Matching Requirements to Techniques
| Requirement | Suitable approach |
|---|---|
| Exact per-decision reason codes for a regulator | Linear or logistic regression, or shallow trees |
| Global validation that the model uses sensible drivers | Global attributions (ML.GLOBAL_EXPLAIN, SHAP summaries) on any model |
| Explain which image region drove a diagnosis | Integrated gradients or XRAI saliency on the vision model |
| Explain an unexpected prediction by showing similar past cases | Example-based explanations or nearest-neighbor retrieval on embeddings |
| Explain a forecast to planners | ARIMA_PLUS decomposition with ML.EXPLAIN_FORECAST |
Interpretability for Generative AI
LLMs don't provide faithful feature attributions, and a model's self-written "reasoning" isn't a guaranteed account of how it produced an answer. For gen AI, interpretability usually means:
- Grounding with citations, so users can check sources (Chapter 4).
- Structured outputs that separate the decision from the supporting evidence.
- Evaluation of grounding and instruction following (Chapter 7).
- Keeping regulated decisions in interpretable models, with the LLM summarizing or drafting only.
Worked Scenario
A lender must give each rejected applicant the top reasons for denial, and model-risk reviewers must approve the model yearly.
- A logistic regression with well-engineered features reaches AUC 0.81. A boosted tree reaches 0.83.
- Adverse-action reasons must be exact and consistent, and the 0.02 AUC gain doesn't offset the review burden. The team picks logistic regression, with reason codes from each feature's contribution (weight × value).
- A separate boosted tree with Tree SHAP explanations ranks marketing offers, where approximate explanations are acceptable.
A bank's model-risk policy requires exact, stable reasons for every credit denial. A logistic regression model reaches AUC 0.80, and a deep neural network reaches AUC 0.81. Which choice best satisfies the requirement?
A team needs local feature attributions for a custom XGBoost model served in a custom container on Agent Platform. Which built-in attribution method matches this non-differentiable model?
Which statement about feature attributions is correct?