While training a churn predictor, the loss on the training split keeps falling, but the loss on a held-out validation split has risen over the last several evaluation rounds. What is the most appropriate stopping rule?
Show answer and explanation
Answer: D
Early stopping watches a held-out validation metric and ends training once it stops improving for a patience window, which prevents the model from memorizing the training data. A fixed pass count ignores what is happening, a near-zero training loss usually signals memorization rather than readiness, and model size has no bearing on overfitting.
Why the other options are wrong
- A. A fixed number of passes can end training too early or too late, and it never checks whether the model has started to memorize the training split.
- B. Near-zero training loss is a common sign of overfitting, so waiting for it means the model has already learned noise in the training data.
- C. Serialized model size reflects architecture choices and says nothing about generalization, so a size cap cannot detect overfitting.