MLS-C01 Question 84
Select 3A data science team is tasked with building a machine learning model to classify images of defective and non-defective products in a manufacturing pipeline. They have collected a dataset containing 5,000 labeled images: 4,800 labeled as 'non-defective' and 200 labeled as 'defective.' The team suspects this class imbalance might affect model performance. What should the team consider to determine if they have sufficient labeled data for the task?
- A
Evaluate whether the dataset represents the real-world distribution of defective and non-defective products.
- B
Perform exploratory data analysis (EDA) to identify if the minority class (defective) has enough examples to train a robust model.
- C
Directly train a deep learning model without assessing the labeled data, as the dataset size is large.
- D
Consider using techniques such as data augmentation or synthetic data generation to address the imbalance in the dataset.
- E
Ensure the dataset has at least 10,000 labeled examples for both classes, as this is the minimum requirement for image classification tasks.
Show answer and explanation
Correct answers: A, B, D
Explanation
When determining if labeled data is sufficient, it is essential to ensure the dataset is representative of real-world conditions and contains enough examples for all classes, especially the minority class. Techniques like exploratory data analysis and addressing class imbalance through augmentation or synthetic data generation are effective strategies. Directly training the model or assuming a fixed dataset size requirement without assessing the dataset's quality and balance is not recommended.
- A. Correct.
This is a correct option. Ensuring the dataset reflects the real-world distribution is important in determining if the labeled data is sufficient and representative for the machine learning model to generalize well in production.
- B. Correct.
This is a correct option. Exploratory data analysis is crucial to understand the distribution of classes and identify if the minority class has sufficient examples to learn meaningful patterns.
- C. Incorrect.
This is incorrect. Directly training a model without assessing the labeled data could lead to poor performance, especially with class imbalance. It is important to evaluate the dataset's sufficiency first.
- D. Correct.
This is a correct option. Data augmentation or synthetic data generation techniques can help address the imbalance if the minority class has insufficient labeled examples.
- E. Incorrect.
This is incorrect. There is no universal minimum dataset size for image classification tasks. Sufficiency depends on factors such as model complexity, class imbalance, and data quality.