When a model performs very well on its training set but poorly on new data, such as a test set, because it has fit the noise and quirks of the training data instead of only the underlying pattern.
Proposed by @Al_Grigor
Card 1 of 259. Answer: When a model performs very well on its training set but poorly on new data, such as a test set, because it has fit the noise and quirks of the training data instead of only the underlying pattern.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
Overfitting is when a model fits its training data so closely, noise and quirks included, that it performs well on that data and noticeably worse on new data from the same source. The difference between training and test performance is the generalization gap, and a large gap is the main symptom.
Example: a sentiment classifier with thousands of word features, trained on 200 reviews, scores 99% on its training data but 71% on held-out reviews. It learned that 'Tuesday' signals a negative review, because three angry reviews happened to mention one. That pattern holds in those 200 reviews and nowhere else; the more parameters per example, the more such accidents a model can fit.
Interviewers then ask how to detect and prevent it: score the model on unseen data, and use more data, fewer features, regularization or early stopping. A subtler form is overfitting the validation set: try enough models against one holdout and the best score is optimistic, which is why Kaggle ranks can fall on the private leaderboard.
