MLA-C01 Question 222
Select 4You are training a machine learning model in Amazon SageMaker to predict customer churn rates for a subscription service. During evaluation, you notice that the training accuracy is very high, but the validation accuracy is significantly lower. Which of the following methods can help you identify and address this issue?
- A
Analyze the learning curves to compare training and validation performance over epochs
- B
Increase the size of the training dataset to improve generalization
- C
Use cross-validation to assess model performance on different subsets of data
- D
Reduce the model complexity by using regularization techniques like L1 or L2
- E
Monitor the training and validation loss to detect divergence or overfitting
Show answer and explanation
Correct answers: A, C, D, E
Explanation
To identify and address overfitting, you should analyze learning curves, monitor training and validation loss, and use techniques like cross-validation as part of your evaluation strategy. Additionally, introducing regularization can help diagnose and mitigate overfitting. Simply increasing the dataset size may help improve generalization but does not directly diagnose the issue.
- A. Correct.
Analyzing learning curves can help you visualize whether the model is overfitting (high training accuracy, low validation accuracy) or underfitting based on the trends in training and validation performance over epochs.
- B. Incorrect.
Increasing the size of the training dataset may improve model performance, but it does not directly help identify overfitting or underfitting. It is more of a corrective action rather than a diagnostic method.
- C. Correct.
Using cross-validation can help you better evaluate the model's performance across different subsets of data, which is useful in identifying overfitting or underfitting.
- D. Correct.
Regularization techniques like L1 or L2 reduce model complexity and can be used to combat overfitting. However, they are also diagnostic tools to experiment with model behavior under different constraints.
- E. Correct.
Monitoring training and validation loss curves is a direct method to detect overfitting (divergence between losses) or underfitting (both losses are high and converge).