Data augmentation creates synthetic training instances by applying transformations to existing examples. The added variation increases effective dataset diversity and helps models learn patterns that remain stable under permitted input changes. section 5.2.2
In imbalanced classification, augmentation provides an alternative to simply repeating minority examples. Repetition can encourage overfitting; slightly modified synthetic instances expose the model to broader variation instead. section 5.2.2
Typical transformations include:
- Rotate, flip, or zoom images so models handle different object orientations and scales. section 5.2.2
- Modify molecular SMILES by adding, removing, or changing functional groups, atoms, or bonds; related changes include bioisosteric replacement. section 5.2.2
Treat augmentation as a regularizer: it reduces memorization and overfitting while improving generalization to unseen instances, real-world variation, and noise. Transformations should preserve the task-relevant signal. section 5.2.2











