Dataset Splitting

Updated 17 Sep 2026

•
What it is
- •
  separatinig data according to its role in developing and evaluating a model
•
Roles
- •
  training set — fit model parameters
- •
  validation set — select models, tune hyperparameters, and support early stopping
- •
  test set — estimate final performance after development decisions are fixed
•
Principles
- •
  partition according to the intended generalization claim
- •
  for unseen-person evaluation, keep each person in only one partition
- •
  for future prediction, respect time order
- •
  fit learned preprocessing only on training data
•
Caution
- •
  ratios such as 70/15/15 are examples, not universal requirements
- •
  repeated development decisions based on test performance turn the test set into development data
•
Related
- •
  Cross-Validation
- •
  Data Leakage (Machine Learning)

Linked references 5