Dataset Splitting

machine-learning statistics

Definition

Dataset Splitting

Dataset splitting is the practice of partitioning a dataset $D$ into disjoint subsets to ensure unbiased model development and evaluation. To avoid data contamination (leakage), information used for preprocessing or parameter estimation must only be derived from the training portion.

Standard partitions include:

1. Training Set ($\approx 80\%$): Utilised for the estimation of model parameters $\theta$.
2. Validation Set ($\approx 10\%$): Utilised for the selection of hyperparameters and model selection.
3. Test Set ($\approx 10\%$): Utilised exclusively for the final evaluation of the model’s generalisation performance. It must remain strictly unseen during all stages of training and tuning.