A Robust Deterministic Data Splitting Method for Model Validation

Random sampling is commonly used for creating a testing set for the purpose of model validation. Deterministic methods exist for data splitting that reduce the variability in the random splitting. Although more consistent conclusions can be obtained using deterministic methods, we point out that they can fail when there are certain structures in the data. In particular, this can happen when the testing points are too close to the training points. In this article, we propose a new deterministic method that maximizes the separation between the training and testing points, but at the same time maintains their distribution. Our simulations using several test functions and datasets show that the proposed method is a robust alternative to random splitting and leave-one-out cross-validation.