Training, Validation, and Test Sets

Imagine you're teaching a student a new subject. You give them study materials and practice problems. How do you know if they've actually learned the concepts, or if they've just memorized the answers to the specific problems you gave them? You wouldn't test them using the exact same practice problems, right? You'd give them a new test with different questions covering the same topics.

Machine learning models face a similar challenge. If we train a model on all the data we have and then test it on that same data, it might perform perfectly simply because it "memorized" the answers. This doesn't tell us how well the model will perform on new, unseen data in the future, which is usually the whole point of building the model. This ability to perform well on new data is called generalization.

To properly evaluate generalization and build reliable models, we split our dataset into separate subsets. The most common approach uses three distinct sets:

Training Set

This is the largest portion of your data, typically comprising 60% to 80% of the total dataset. The machine learning algorithm uses this data to actually learn the patterns and relationships between the input features and the output labels (in supervised learning). Think of this as the model's "textbook" and "homework problems". The model iteratively adjusts its internal parameters by looking at the examples in the training set, trying to minimize the errors it makes on this data.

Validation Set (or Development Set)

This is a smaller, independent chunk of data, perhaps 10% to 20% of the total. The model does not directly train on this data. Instead, the validation set serves two main purposes:

Hyperparameter Tuning: As we learned previously, hyperparameters are settings we choose before training begins (unlike parameters, which the model learns during training). Examples might include choosing the complexity of the model or setting a learning rate for the training algorithm. We often try several different combinations of hyperparameters, train separate models using the training set for each combination, and then evaluate each trained model's performance on the validation set. The hyperparameter settings that yield the best performance on the validation set are typically chosen for the final model.
Model Selection & Overfitting Check: We can also use the validation set to compare different types of models (e.g., should we use Linear Regression or KNN for this problem?). Furthermore, by monitoring the model's performance on the validation set during the training process, we can detect if the model is starting to perform extremely well on the training data but poorly on the validation data. This divergence often indicates overfitting, where the model has memorized the training data too closely, including its noise, and is losing its ability to generalize. The validation set acts as an early warning system.

Think of the validation set as the "quizzes" or "practice exams" used during the learning process. It helps guide the study strategy (hyperparameter tuning) and checks for genuine understanding (generalization) without using the final exam questions.

Test Set

This final subset, also typically 10% to 20% of the data, is held back and used only once after all training and hyperparameter tuning are complete. The model has never seen this data before during its development.

"The sole purpose of the test set is to provide an unbiased, final evaluation of the chosen model's performance. It simulates how the model is expected to perform on completely new data."

It's absolutely essential that the test set is not used for any training or tuning decisions. If you evaluate on the test set, then go back and tweak your model or hyperparameters based on those results, and then evaluate again, you have effectively "contaminated" the test set. It no longer provides an unbiased estimate because your decisions were influenced by its specific data. Think of it like the final, official exam. You take it once, and that score represents your knowledge. You don't get to retake it repeatedly after seeing the questions.

Visualizing the Split

We can visualize this division of data:

A typical workflow involves splitting the initial dataset into training, validation, and test sets. The model learns from the training set, is tuned using the validation set, and finally evaluated on the test set.

Common Practices

The data is usually shuffled randomly before splitting to ensure that each set is representative of the overall distribution of the data. The exact percentages (e.g., 70/15/15, 80/10/10) can vary depending on the size of the dataset and the specific problem. For very large datasets, smaller percentages might be sufficient for validation and testing. Sometimes, especially with simpler workflows or smaller datasets, people might only split into training and testing sets (e.g., 80/20), perhaps using techniques like cross-validation on the training set to perform hyperparameter tuning and model selection (we'll likely encounter cross-validation later). However, the three-set split (train/validate/test) is a foundation and widely accepted practice, especially when developing more complex models.

In summary, splitting data into training, validation, and test sets is a fundamental technique in machine learning. It allows us to train models, tune them effectively, and get a reliable estimate of how well they will perform on new data, which is essential for building trustworthy and useful applications. This practice directly helps in diagnosing and preventing common issues like overfitting, a topic we will explore shortly.

Was this section helpful?

References

Pattern Recognition and Machine Learning, Christopher M. Bishop, 2006 (Springer) DOI: 10.1007/978-0-387-45528-0 - This classic textbook covers the fundamental principles of machine learning, including model evaluation, generalization, and the use of validation sets for model selection and hyperparameter tuning.
Deep Learning, Ian Goodfellow, Yoshua Bengio, and Aaron Courville, 2016 (MIT Press) DOI: 10.7551/mitpress/9780262035613.001.0001 - Chapter 5 ("Machine Learning Basics") offers an authoritative explanation of generalization, overfitting, underfitting, and the role of training, validation, and test sets in deep learning contexts, applicable to general machine learning.
CS229 Lecture Notes: Supervised Learning (part 2), Andrew Ng, 2019 Stanford University CS229 - These widely used lecture notes from a foundational machine learning course explain data splitting for model evaluation and selection in an accessible manner.
Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, Aurélien Géron, 2022 (O'Reilly Media) - This practical guide explains the importance of train/validation/test sets for building and evaluating machine learning models, with clear examples of how to implement the split. (3rd edition)