Traditionally, the validation set is crucial in the ML workflow, providing a dataset to optimize model parameters without introducing bias to the final model evaluation. In practice, it's employed during model development to iteratively adjust hyperparameters and architecture decisions. The effectiveness of the validation set lies in preventing overfitting to the training data, facilitating a more generalizable model. In complex models, especially deep learning scenarios, the validation set plays a pivotal role in stopping training early when performance no longer improves, known as early stopping, enhancing both the efficiency and performance. The use of a validation set, separated from both the training and testing data, is often referred to in the cross-validation process, where multiple splits and rotations are used for robustness before the final model evaluation.
The term "validation set" emerged as data-centric ML methods gained popularity in the late 1980s and early 1990s, although its formal distinction from training and testing sets became standard during the rise of more sophisticated model validation techniques in the late 1990s.
Key contributors to the development of the validation set concept include practitioners in the early stages of ML development, such as advocates of cross-validation techniques. While no single individual is credited with creating the validation set framework, its use became prominent through the work of researchers who focused on improving model generalizability, contributing to its standard practice in modern ML pipelines.