Katseandmed

Katseandmed, Estonian for test data, denote the portion of a dataset reserved for evaluating a model, algorithm, or hypothesis after training. Unlike training data, katseandmed are not used to fit the model; they provide an unbiased estimate of performance in deployment conditions.

Typically, data projects split data into training, validation, and test sets. The training set fits the model, the validation set guides hyperparameter tuning and model selection, and the katseandmed test set is used only for final assessment. In time-sensitive or sequential data, splits should preserve temporal order to avoid leakage.

To be effective, katseandmed should reflect the deployment context: representative of the target population or input distribution, of sufficient size to yield stable metrics, and free from overlap with training data. Data leakage, such as including features derived from or correlated with labels in the test set, must be avoided. Reproducibility is facilitated by fixed random seeds or documented splitting procedures.

Common evaluation metrics on katseandmed depend on the task: classification uses accuracy, precision, recall, F1, and ROC AUC; regression uses RMSE, MAE; ranking or multi-label tasks have respective metrics. Reported results typically accompany confidence intervals or multiple seeds to convey uncertainty.

Ethical and methodological considerations include privacy-preserving handling, de-identification, and attention to biases in the test set. Proper documentation, such as a data sheet describing data provenance and collection methods, supports transparency and comparability.