A credit-scoring model's production inputs have gradually shifted away from the sample it was trained on, so its predictions are becoming less reliable. What is this phenomenon called?
Show answer and explanation
Answer: A
Data drift occurs when the distribution of inputs seen in production differs from the data used for training, so a model that was accurate at launch gradually loses reliability. A related problem, concept drift, happens when the relationship between inputs and the correct output changes. Overfitting is a training-time problem in which a model memorizes noise in its training data, feature explosion describes an unmanageably large feature set, and gradient vanishing is an optimization failure in deep networks. Monitoring input distributions detects data drift early.
Why the other options are wrong
- B. Overfitting shows up during training as strong training results and weak generalization, not as a shift in production inputs.
- C. Feature explosion refers to an unmanageably large feature set, which is a design issue rather than a change in live data.
- D. Vanishing gradients are an optimization problem during training in deep networks, not a change in production inputs.
Exam tip. Know the drift types: data drift changes inputs, concept drift changes the input-output relationship, and prediction drift changes outputs.