classweighted

Classweighted is a technique in supervised learning used to address class imbalance by assigning different importance to classes during model training. By giving higher weight to underrepresented classes, the method aims to improve the model's performance on those classes without requiring data resampling.

Weights are typically derived from class frequencies. Common schemes include w_c = N / (K * n_c) and w_c = 1 / n_c, where N is the total number of samples, K is the number of classes, and n_c is the count of samples in class c. Some libraries offer a “balanced” option that computes these weights automatically. In multi-class problems the weight for each class is independent and is used to scale the loss contributed by samples of that class.

Implementation details vary by framework. In neural networks, a per-class weight vector is passed to loss functions such as cross-entropy, for example in PyTorch via CrossEntropyLoss(weight=...), which multiplies the loss for each sample by the weight of its true class. In scikit-learn, class weights can be provided to estimators or computed with compute_class_weight. In Keras and TensorFlow, class_weight is supplied to fit(), or sample_weight can be used for finer control.

Considerations and limitations include the fact that weighting alters the optimization objective and can improve minority-class recall at the expense of overall accuracy. Overemphasis on rare classes can lead to overfitting. Proper evaluation with metrics that reflect class balance (precision, recall, F1-score, balanced accuracy) is recommended. Alternatives or complements include data resampling (oversampling minority or undersampling majority) and loss variants such as focal loss.

See also: class imbalance, resampling, focal loss, sample_weight.