MLA-C01 Question 98
Select 4An e-commerce company is building a machine learning model to recommend products to users. Their dataset contains numeric features (e.g., user age, purchase count), text features (e.g., product reviews), and image features (e.g., product photos). However, the dataset is highly imbalanced, with very few examples of certain product categories. What strategies can the company use to address class imbalance (CI) across the numeric, text, and image datasets?
- A
Generate synthetic data for minority classes using techniques like SMOTE for numeric features and GANs for image features.
- B
Use oversampling techniques such as duplicating minority class samples to balance the dataset.
- C
Apply dimensionality reduction techniques like PCA to remove irrelevant features from the text dataset.
- D
Use weighted loss functions during model training to give higher importance to the minority class.
- E
Manually collect additional data for the minority class to improve representation.
Show answer and explanation
Correct answers: A, B, D, E
Explanation
Class imbalance can significantly affect the performance of machine learning models, particularly when working with diverse data types such as numeric, text, and image data. Strategies such as synthetic data generation (e.g., SMOTE for numeric data, GANs for image data), oversampling, weighted loss functions, and manual data collection are effective ways to address imbalances. Dimensionality reduction like PCA, while useful for feature selection, does not specifically target class imbalance.
- A. Correct.
Correct: Synthetic data generation techniques like SMOTE (for numeric data) and GANs (for image data) can create new samples for the minority class, helping to address class imbalance.
- B. Correct.
Correct: Oversampling methods, such as duplicating minority class samples, are straightforward and effective for balancing imbalanced datasets.
- C. Incorrect.
Incorrect: PCA and dimensionality reduction are used to reduce the number of features, not specifically to address class imbalance in datasets.
- D. Correct.
Correct: Weighted loss functions can directly mitigate imbalances by ensuring the minority class has a larger impact on the model's optimization.
- E. Correct.
Correct: Collecting additional data for the minority class is a valid and effective strategy to enhance representation and reduce class imbalance.