MLS-C01 Question 86
Select 2You are building a machine learning model for predicting fraudulent transactions in an e-commerce application. After deploying the model, you notice that it is producing biased results, with certain demographic groups being unfairly flagged as fraudulent more often than others. Which of the following strategies should you adopt to mitigate this bias effectively?
- A
Re-train the model using a dataset that is balanced across all demographic groups.
- B
Apply feature engineering to remove demographic information from the dataset before training the model.
- C
Use explainability tools such as Amazon SageMaker Clarify to identify and address bias in the data and model.
- D
Increase the learning rate of the model to ensure faster convergence on unbiased results.
- E
Implement threshold tuning to adjust the decision threshold for different demographic groups.
Show answer and explanation
Correct answers: A, C
Explanation
Bias in machine learning models can arise due to imbalanced training data or the presence of features that unfairly influence predictions. Re-training the model on a balanced dataset ensures that all groups are represented fairly during training, while using tools like Amazon SageMaker Clarify helps uncover and address bias in both the data and the model. These strategies are effective and widely recommended for mitigating bias in machine learning workflows.
- A. Correct.
Re-training the model with a balanced dataset can help reduce bias by ensuring that all demographic groups are fairly represented in the training data, which is a key strategy for mitigating bias.
- B. Incorrect.
Removing demographic information might seem like a solution, but this can lead to unintentional proxy bias, where other features indirectly correlate with the removed demographic information and perpetuate the bias.
- C. Correct.
Using explainability tools like Amazon SageMaker Clarify allows you to identify and quantify biases in both the dataset and the model predictions, enabling targeted mitigation strategies.
- D. Incorrect.
Increasing the learning rate does not directly address the issue of bias and could lead to unstable training or convergence issues.
- E. Incorrect.
Threshold tuning for different demographic groups may introduce additional complexity and ethical concerns, and it is not a recommended first step for mitigating bias in models.