MLA-C01 Question 178
Select 3You are working on a machine learning model deployed on Amazon SageMaker, and the model's accuracy on a validation dataset is lower than expected. After analyzing the data, you notice that the dataset is imbalanced, and the model performs poorly on the minority class. Which actions can you take to improve your model's performance?
- A
Apply oversampling techniques to balance the dataset
- B
Use a different optimizer, such as Adam instead of SGD
- C
Introduce class weights during model training
- D
Collect more data for the minority class
- E
Switch to a simpler model architecture with fewer parameters
Show answer and explanation
Correct answers: A, C, D
Explanation
When dealing with imbalanced datasets, techniques such as oversampling the minority class, introducing class weights, or collecting more data for the minority class are effective in improving model performance. These approaches focus on addressing the imbalance issue directly, enabling the model to better learn from underrepresented data. Changing the optimizer or simplifying the model architecture does not target the imbalance problem and is less relevant in this scenario.
- A. Correct.
Applying oversampling techniques, such as SMOTE, can help balance the dataset by synthetically increasing the representation of the minority class, which improves the model's ability to learn from it.
- B. Incorrect.
Changing the optimizer affects how the model converges during training but does not directly address the class imbalance issue. Therefore, this option is not the most effective for this scenario.
- C. Correct.
Introducing class weights can help the model prioritize learning from the minority class by giving it higher importance during the loss computation, which can improve performance on imbalanced datasets.
- D. Correct.
Collecting more data for the minority class directly addresses the imbalance issue, providing the model with more examples to learn from, which improves its performance.
- E. Incorrect.
Switching to a simpler model architecture could reduce overfitting but does not address the root cause of the imbalance in the dataset and is unlikely to improve minority class performance.