Klassenimbalance

Klassenimbalance, in supervised machine learning often simply called class imbalance, describes a situation where one class occurs much more frequently than others. In binary problems, the positive or minority class is underrepresented; in multi-class problems, several classes may be rare. Such distributions reflect many real-world tasks, including fraud detection, medical diagnosis, and anomaly screening.

The main consequence of imbalanced data is that models tend to favor the majority class. This can lead to high overall accuracy while the minority class is poorly detected, producing biased decision boundaries and poorly calibrated probability estimates. Evaluation based on overall accuracy can be misleading, so it is important to examine metrics that reflect performance on the minority class and overall balance.

Commonly used measurements include the class distribution ratio, recall, precision, F1-score, and the geometric mean of class-wise accuracies. For highly skewed problems, area under the precision–recall curve (AUPRC) and balanced accuracy are often more informative than AUC-ROC, which can mask minority-class performance.

Several strategies exist to address klassenimbalance. Data-level approaches include oversampling the minority class (random oversampling, SMOTE, ADASYN) and undersampling the majority class, sometimes with data-cleaning steps like Tomek links. Algorithmic approaches adjust the learning objective or model parameters by using class weights or cost-sensitive loss functions, or by tuning decision thresholds. Ensemble methods, such as balanced random forests or boosting variants, can also help. Cross-validation should be stratified, and evaluation should rely on metrics that reflect minority-class performance to ensure robust conclusions.