Core Concepts
Early stopping of cross-validation during model selection can make the process more effective by allowing model selection to converge faster and explore the search space more exhaustively, while also obtaining better overall performance.
Abstract
The content discusses the use of early stopping methods for cross-validation during model selection in automated machine learning (AutoML) systems. The authors aim to make model selection with cross-validation more effective for AutoML.
The key highlights and insights are:
The authors present two simple-to-understand and easy-to-implement early stopping methods, Aggressive and Forgiving, and compare them to the baseline of not using early stopping.
Experiments are conducted on 36 classification datasets, using random search and Bayesian optimization as the model selection strategies, and considering 3-, 5-, and 10-fold cross-validation scenarios, as well as repeated cross-validation.
The results show that Forgiving early stopping consistently allows model selection with random search to converge faster (by 214% on average) and explore the search space more exhaustively (by 167% more configurations on average) within a one-hour time budget, while also obtaining better overall performance.
Aggressive early stopping can also lead to speedups, but fails to match or improve over the best performance found by the baseline in roughly half of the datasets, likely due to overly aggressive early stopping of configurations that would otherwise yield good performance.
The authors also investigate the impact of early stopping on Bayesian optimization, finding that Forgiving can lead to better overall performance, although to a lesser extent than for random search.
The positive effect of early stopping cross-validation is also observed for repeated cross-validation, with Aggressive even outperforming Forgiving in the 2-repeated 10-fold case.
Overall, the study demonstrates the advantages of early stopping cross-validation for model selection in AutoML and provides a simple-to-understand, easy-to-implement, and well-performing method for doing so.
Stats
Fitting and validating the configuration of an MLP was empirically ∼10.5× more expensive when going from a 90/10 holdout validation to 10-fold cross-validation on the okcupid-stem dataset.
On average, Forgiving early stopping allowed model selection with random search to converge 214% faster than the baseline of no early stopping.
On average, Forgiving early stopping allowed model selection with random search to explore 167% more configurations within the one-hour time budget compared to the baseline.
Quotes
"We aim to make model selection with cross-validation more effective for AutoML."
"Our study shows that a simple-to-understand and easy-to-implement method for early stopping cross-validation (1) consistently allows model selection with random search to converge faster, in ∼94% of all datasets, on average by 214%; and (2) explore the search space more exhaustively by considering +167% configurations on average within the time budget of one hour; while also (3) obtaining better overall performance."