Data Split

Data Split is essential for sound and effective model training and validation. It lets users divide a dataset into training, validation, and test sets using predefined ratios or strategies such as chronological splitting, random splitting, or feature-based stratified sampling. It simplifies data preparation, supports fair evaluation, helps prevent overfitting, and provides stronger evidence of model generalization.

General Functions

Our Data Split functions support multiple strategies, including random splitting and feature-based stratified sampling, to meet the needs of different problems. They reduce preprocessing complexity and improve the accuracy of model evaluation and the credibility of research results.

— Splits data into training, test, and validation sets according to frac; the default ratio is 7:2:1.

— For small datasets, k-fold cross-validation can be used to split the dataset.

Prediction

For Prediction tasks, Data Split primarily supports time-series splitting to meet the needs of short- and long-horizon prediction.

— Raw datasets have different sampling rates, so frame counts are used as inputs: frame_input frames are used to predict frame_output frames.