Rachel (General Counsel) has regulatory concerns.

Rachel: "If we deny someone's transaction, we need to explain WHY. Can your model do that?"

Your model's explainability: [based on Q1 choice]

What do you do?

Current tokens: [accumulated]

Explainability based on Q1:

Logistic Regression: HIGH (coefficients are clear)
Decision Tree: HIGH (can show decision path)
Random Forest: MEDIUM (feature importance available)
XGBoost: LOW (black box)
Neural Network: VERY LOW (complete black box)

Question

Rachel (General Counsel) has regulatory concerns.

Rachel: "If we deny someone's transaction, we need to explain WHY. Can your model do that?"

Your model's explainability: [based on Q1 choice]

What do you do?

Current tokens: [accumulated]

Explainability based on Q1:

Logistic Regression: HIGH (coefficients are clear)
Decision Tree: HIGH (can show decision path)
Random Forest: MEDIUM (feature importance available)
XGBoost: LOW (black box)
Neural Network: VERY LOW (complete black box)

BlackTom AI · Accepted Answer

SHAP (Shapley Additive Explanations) provides model-agnostic, per-decision feature attributions, so adding a SHAP explainability layer is the appropriate action.

1) Relevant knowledge points and concepts
- Interpretability levels: Logistic Regression and Decision Trees are inherently interpretable (coefficients, decision paths). Random Forests give global feature importances but weak per-instance explanations. XGBoost and Neural Networks are effectively black boxes.  
- SHAP uses Shapley values from cooperative game theory to assign each feature a contribution to a single prediction; it yields additive explanations that sum to (model output) = (baseline) + sum(feature contributions).  
- There are optimized SHAP implementations (TreeSHAP) for tree ensembles and model-agnostic variants (KernelSHAP) that work with any model.

2) Step-by-step reasoning process
- Regulatory need: Rachel requires a clear WHY for each denied transaction (local explanation).  
- Compare options: Interpretable models (logistic, tree) can explain directly but switching models may be infeasible or degrade performance. Random Forests give partial info; XGBoost/Neural Nets are opaque.  
- Best practical solution: keep the chosen predictive model and add SHAP as an explanation layer to produce standardized, human-readable reasons for each decision.

3) Example calculation (intermediate steps)
- Suppose baseline fraud risk = 0.20. Model predicts 0.80 (deny). SHAP yields contributions: income: −0.10, unusual_txn_count: +0.40, device_risk: +0.30.  
- Sum: 0.20 + (−0.10) + 0.40 + 0.30 = 0.80.  
- Presentable explanation: “Transaction denied because unusual transaction count (+0.40) and device risk (+0.30) increased risk above baseline; low income reduced risk slightly (−0.10).”

4) Why this answer is correct
- SHAP gives per-transaction feature contributions, is actionable for regulators, and works across model types—exactly meeting Rachel’s requirement.

5) Why other options are inferior
- Logistic Regression / Decision Tree: already interpretable but may not be the deployed model or may lose accuracy if swapped.  
- Random Forest: only medium—global importances don’t explain individual decisions well.  
- XGBoost/Neural Net: low/very low interpretability without an add-on like SHAP.

Conclusion: Adding a SHAP explainability layer (Cost: 2) is the correct, model-agnostic way to give clear per-decision explanations; this matches the provided answer.

类似问题

An image-recognition model labels a medical image as suspicious and highlights the regions that most influenced the classification. What is the most appropriate description of how the highlighting functions as an explanation for patients and doctors?

A hospital deploys an AI sepsis early-warning system that gives emergency department physicians a risk score and brief rationale for each patient. Which design choice would most directly support appropriately calibrated trust by the intended users?

Which of the following best describes why explainability is a subject of ethical significance in "high-stakes" AI systems?

What does London mean wen he describes deep learning AI's reasoning as "atheoretical, associationist, and opaque"? (Check BOTH that apply)

Which of the following best describes a contrastive explanation in AI?

The words interpretable and explainable are closely related. However, in machine learning they refer to slightly different aspects of model behaviour. Choose the most appropriate definition below.

What is the purpose of explainable AI?

更多留学生实用工具

智能学习助手

风格化写作助手

论文查重助手

文献引用助手

课堂转译助手

课堂笔记助手

Quiz搜索助手

学校历年真题

智能刷题助手

智能匹配练习

希望你的学习变得更简单