Answer: Partial dependence plots (PDPs)

## Explanation To meet transparency and explainability requirements for stakeholders in a quarterly demand forecasting context, the AI practitioner should include **Partial Dependence Plots (PDPs)** in the report. Here's the detailed reasoning: ### Why Partial Dependence Plots (PDPs) are Optimal: 1. **Visual Explanation of Feature Impact**: PDPs graphically show how changes in specific input features (e.g., marketing budget, seasonal factors, economic indicators) affect the model's predictions while holding other features constant. This allows stakeholders to understand which factors most influence demand forecasts without needing technical ML knowledge. 2. **Addresses Stakeholder Questions**: In demand forecasting, stakeholders often ask questions like "What happens to predicted demand if we increase our advertising spend by 20%?" or "How sensitive are our forecasts to changes in raw material costs?" PDPs directly answer these causal questions by illustrating the relationship between individual features and model outputs. 3. **Model Transparency**: PDPs reveal whether the model's behavior aligns with business intuition. For example, if a PDP shows that increased marketing expenditure leads to higher predicted demand (as expected), this builds confidence in the model. Conversely, if it shows counterintuitive relationships, it flags potential issues that need investigation. 4. **Interpretability for Non-Technical Audiences**: Unlike complex model internals or convergence metrics, PDPs are intuitive visualizations that business stakeholders can easily interpret. This makes them ideal for reports aimed at executives, operations managers, and other non-technical decision-makers. 5. **AWS Best Practices Alignment**: AWS emphasizes model interpretability tools like PDPs for explaining ML predictions to stakeholders. In the context of AWS AI/ML services, such visual explanations are recommended for building trust in automated decision-making systems. ### Why Other Options Are Less Suitable: - **A: Code for model training**: While transparency about methodology is important, sharing raw code doesn't provide meaningful explainability to non-technical stakeholders. Code is implementation detail rather than business insight. - **C: Sample data for training**: Sample data shows what the model was trained on but doesn't explain how the model makes predictions or what drives those predictions. - **D: Model convergence tables**: These are technical metrics useful for data scientists during model development but don't provide business-relevant explanations about how features affect forecasts. ### Additional Considerations for the Report: While PDPs are the core component for explainability, a comprehensive report might also include: - **Business context** explaining how the forecasts will be used - **Model performance metrics** relevant to business outcomes - **Limitations and assumptions** of the forecasting approach - **Recommendations** based on model insights However, among the given options, PDPs are specifically designed to provide the transparency and explainability stakeholders need to understand and trust the demand forecasting models.