What is dataset leakage and how can it be prevented?

• A. Leakage occurs when information from the target or test set leaks into training; prevent with proper train/test splits, time-sequenced data, and clean pipelines.
• B. Leakage occurs when information from the target or test set leaks into training; prevent with adding more noise to the data.
• C. Leakage occurs when the model overfits; prevent with larger networks.
• D. Leakage occurs when you shuffle training data.

Answer: (A)

Dataset leakage means information that should only be known after making a prediction ends up in the training data, so the model learns from data that would not be available in real use. This makes evaluation look unrealistically good because the model has effectively seen clues about the target during training.

The correct approach focuses on preventing that leakage. Use train-test splits that reflect how the model will be used in practice, ensuring the test set remains truly unseen. For data with a time component, apply time-sequenced splits so the model trains only on past information and is evaluated on future data. Build clean preprocessing pipelines so that steps like scaling, encoding, or imputing are learned from the training data only and then applied to the test data, avoiding leakage from statistics calculated on the full dataset. Also be careful not to include the target label in features or derive features in ways that reveal the answer.

Adding more noise doesn't address leakage and can just degrade performance. Overfitting is a symptom of the model fitting to training data too closely and is not a leakage prevention method. Shuffling data can be inappropriate for time-ordered data and doesn't directly prevent leakage.