Which data preprocessing step reduces the risk of bias and leakage in CPMAI data handling?

Which data preprocessing step reduces the risk of bias and leakage in CPMAI data handling?

• A. Proper data splitting to avoid leakage, normalization, and careful handling of imbalanced data.
• B. Randomly removing features to reduce model capacity.
• C. Only normalizing data without splitting.
• D. Increasing dataset size without proper splitting.

Answer: (A)

The key idea is preventing data leakage and bias through careful preprocessing. Properly splitting the data into training (and validation) and test sets first ensures that information from the test set never informs the model during training. Then, normalization should be fit on the training data alone and applied to both training and test data, so the scaling parameters aren’t influenced by the test set. Additionally, addressing imbalanced data helps avoid bias toward the majority class, supporting more fair, generalizable results. Other approaches don’t address these risks effectively: removing features randomly reduces model capacity but doesn’t tackle leakage or bias from data handling; normalizing without any split can let information from the whole dataset influence the model, introducing leakage; simply increasing dataset size without proper splitting doesn’t remove leakage and may still lead to biased evaluation.

The key idea is preventing data leakage and bias through careful preprocessing. Properly splitting the data into training (and validation) and test sets first ensures that information from the test set never informs the model during training. Then, normalization should be fit on the training data alone and applied to both training and test data, so the scaling parameters aren’t influenced by the test set. Additionally, addressing imbalanced data helps avoid bias toward the majority class, supporting more fair, generalizable results.

Other approaches don’t address these risks effectively: removing features randomly reduces model capacity but doesn’t tackle leakage or bias from data handling; normalizing without any split can let information from the whole dataset influence the model, introducing leakage; simply increasing dataset size without proper splitting doesn’t remove leakage and may still lead to biased evaluation.