MLA-C01 Question 93
Single answerYou are building a machine learning model to classify images of fruits into categories such as apples, bananas, and oranges. During the exploratory data analysis phase, you notice that 70% of the images are labeled as 'apples', while 'bananas' and 'oranges' each constitute 15% of the dataset. What pre-training bias metric would be most appropriate to quantify this imbalance in the dataset?
- A
Class Imbalance (CI)
- B
Difference in Proportions of Labels (DPL)
- C
Mean Absolute Error (MAE)
- D
F1 Score
Show answer and explanation
Correct answer: A
Explanation
Class Imbalance (CI) is a critical pre-training bias metric that helps identify and quantify unequal distribution of classes in a dataset. In this scenario, the dataset has a significant skew in the representation of labels (70% apples, 15% bananas, 15% oranges), which can negatively impact the model's performance. Identifying this issue early allows for corrective actions such as re-sampling or data augmentation.
- A. Correct.
Class Imbalance (CI) is the most appropriate metric to quantify the disproportionate representation of labels in the dataset. It directly measures the unequal distribution of classes.
- B. Incorrect.
Difference in Proportions of Labels (DPL) is typically used to evaluate fairness or bias across different groups after training, not to measure pre-training class distribution.
- C. Incorrect.
Mean Absolute Error (MAE) is a regression metric for assessing prediction errors and is not relevant for analyzing class distribution or pre-training bias.
- D. Incorrect.
F1 Score is a classification metric used to measure the balance between precision and recall in model evaluation, not a pre-training bias metric.