Imbalanced Data
A training set is imbalanced when some classes, or some ranges of a continuous target, are far more common than others. Models trained on such data drift toward predicting the majority, and aggregate metrics like accuracy or mean absolute error can look good while the minority cases are predicted badly. For classification, common remedies reweight or resample the classes and report metrics that are robust to imbalance, such as the F1 score and the Matthews correlation coefficient. For continuous targets with a large point mass at zero, statistics offers two-part (hurdle) models that first decide whether the outcome is zero and then model its size only when it is not. Materials datasets are often imbalanced because databases contain many more examples of common chemistries or states. See also graph neural networks and magnetic moments.
Zero-inflated magnetic moments (2024)
Sanspeur and Kitchin showed that recasting magnetic ordering classification as per-atom moment regression turns class imbalance into a zero-inflated continuous distribution: about 90% of atoms in their antiferromagnetic Materials Project set have moments near 0 µB, and a naive regressor simply predicts the mode. A two-part scheme, with a class-weighted GNN classifier (F1 and MCC about 91%) followed by regression on the embeddings of atoms above a 0.1 µB cutoff, removed the bias and gave a mean absolute error of 0.1 µB (Sanspeur 2024). Summary: sanspeur2024.
Error imbalance across material classes in OC20 (2024)
Imbalance can also appear in the errors rather than the labels. In the OC20 catalysis dataset, machine learning potential errors are highest for nonmetals and halides and lowest for intermetallics. Abdelmaqsoud and coworkers traced much of this to adsorbate-induced surface reconstructions, which are more frequent in nonmetals and halides and make the adsorption energy label inconsistent; removing them from validation reduced the largest class errors the most, and total energy models showed a more uniform error distribution across classes (Abdelmaqsoud et al. 2024). Summary: abdelmaqsoud2024a. See also Adsorption energy referencing.