AIF-C01 Question 192
Single answerYou are building a machine learning model to detect fraudulent transactions in a financial system. The stakeholders require high accuracy to minimize false positives and false negatives, but they also need the system to provide clear explanations for why a transaction was flagged as fraudulent to comply with regulatory requirements. Which trade-off should you consider when designing this model?
- A
Prioritize a highly interpretable model, such as a decision tree, even if it sacrifices some performance.
- B
Choose a black-box model, like a deep neural network, to ensure the highest accuracy and disregard interpretability requirements.
- C
Implement an ensemble of models to achieve both interpretability and performance without any compromise.
- D
Use a high-performing model, such as a gradient-boosted tree, and incorporate explainability tools like SHAP or LIME.
Show answer and explanation
Correct answer: D
Explanation
The correct answer is to use a high-performing model, such as a gradient-boosted tree, and enhance it with explainability tools like SHAP or LIME. This approach allows you to meet the stakeholders' need for both model accuracy and interpretability. Purely interpretable models may not provide sufficient accuracy, while black-box models may fail to meet regulatory requirements. Tools like SHAP or LIME help bridge this gap by providing explanations for predictions without compromising too much on performance.
- A. Incorrect.
Decision trees are highly interpretable, but they may not achieve the level of performance required for the accuracy demanded by the stakeholders.
- B. Incorrect.
Deep neural networks may provide high accuracy, but they lack interpretability, which is necessary for regulatory compliance.
- C. Incorrect.
While ensembles can improve performance, they often lack interpretability, and achieving both without any compromise is unrealistic in most cases.
- D. Correct.
Using a high-performing model, like a gradient-boosted tree, and augmenting it with explainability tools strikes a balance between accuracy and interpretability, addressing both stakeholder requirements.