MLS-C01 Question 367
Select 3A data scientist is building a machine learning model to predict whether an email is spam or not. During model training, they notice that their model achieves an accuracy of 95% but performs poorly on the validation set, with a high false negative rate. Which of the following approaches could help address this issue?
- A
Collect more labeled data to better represent the distribution of spam emails.
- B
Use a different metric, such as F1-score, to better evaluate the model's performance.
- C
Increase the complexity of the model by adding additional layers or parameters.
- D
Apply regularization techniques to reduce overfitting.
- E
Lower the threshold for classifying an email as spam.
Show answer and explanation
Correct answers: A, B, D
Explanation
The key issues in this scenario are overfitting and evaluating the model with an inappropriate metric for an imbalanced dataset. Collecting more labeled data helps the model generalize better. Switching to F1-score ensures the model is evaluated on its ability to balance false positives and false negatives. Regularization reduces overfitting by penalizing overly complex models. While lowering the threshold might help reduce false negatives, it doesn't address the fundamental issues in the training process.
- A. Correct.
Collecting more labeled data could improve the model's understanding of the spam email distribution, potentially addressing the high false negative rate and improving generalization.
- B. Correct.
Switching to a metric like F1-score, which balances precision and recall, is more appropriate for imbalanced datasets where false negatives are critical.
- C. Incorrect.
Increasing model complexity could exacerbate overfitting, worsening performance on the validation set.
- D. Correct.
Applying regularization techniques (e.g., L1 or L2 regularization) can help reduce overfitting, improving the model's generalization to the validation set.
- E. Incorrect.
Lowering the classification threshold might reduce false negatives, but it does not directly address the underlying overfitting or evaluation metric issues.