Data leakage

1 · What it is

Data leakage happens when a model is built using information it will not have when it makes real predictions. The model can then take a shortcut: it uses the leaked detail instead of learning the real pattern. Testing rewards the shortcut, so the score comes out too optimistic, and the model does worse on genuinely new data once it is in use.

The first kind, called label leakage, is a feature that stands in for the answer. Google’s crash course describes a model meant to spot cancer in new hospital patients. It did brilliantly in testing and badly in real use. The reason was the hospital name: patients were sent to a specialist hospital only after diagnosis, so at prediction time most had no hospital yet. A retail version is predicting a day’s revenue from that day’s customer count, which nobody knows until the day’s sales are done.

The second kind lets test data shape training. If a step such as scaling, filling in missing values or choosing features is learned from the whole dataset before splitting, the test set has already influenced the model; the feature scaling entry shows one case. In scikit-learn’s demonstration the labels were random, so about 50 per cent accuracy was expected. Choosing features before splitting pushed accuracy far above chance; splitting first brought it back down. Duplicate rows across the split and shuffled time-ordered data, which lets a model train on the future, leak the same way. For language models the version to watch is benchmark contamination, where test questions turn up in web-crawled training data.

The damage is real. A 2023 survey found leakage in at least 294 papers across 17 fields. In civil war prediction, the four papers claiming complex models beat logistic regression all had leaks. Once fixed, the complex models were no better. A suspiciously near-perfect score is a reason to look for a leak.