A data scientist is developing a classification model to detect fraudulent transactions. The dataset is split into training and test sets. The data scientist repeatedly tunes the model's hyperparameters and evaluates performance on the test set until the test accuracy reaches 95%. However, when the model is deployed on new, unseen data, its accuracy drops to 70%. Which concept best explains this performance degradation?

Overfitting to the training data

Overfitting to training data would cause high training accuracy but low test accuracy, but the tuning process described uses the test set directly, not the training set.

Data leakage from the training set to the test set

Data leakage typically occurs when information from outside the training set influences the model during training, e.g., future data used to predict the past. This scenario does not describe such leakage.

Underfitting the training data

Underfitting would result in poor performance on both training and test sets, which is not the case here as the model achieved 95% test accuracy.

What does this AI-900 question test?

Read the scenario before looking for a memorised answer.

What is the correct answer to this question?

The correct answer is: Overfitting to the test set — This scenario describes overfitting to the test set (also known as 'test set leaking' or 'data snooping'). By repeatedly evaluating on the test set and adjusting hyperparameters to maximize test accuracy, the data scientist inadvertently allowed the test set to influence the model selection. This causes the model to become overly specialized to the test data and fail to generalize to new data. The proper practice is to use a separate validation set for tuning and only use the test set for final evaluation.

What should I do if I get this AI-900 question wrong?

Then try more questions from the same exam bank and focus on understanding why the wrong options are tempting.