What formula does scikit-learn use for balanced class weights?

sklearn computes w_c = n_samples / (n_classes * n_samples_c) for each class c. This ensures that each class contributes equally to the total loss regardless of how many samples it has.

When should I use class weights instead of oversampling?

Class weights are preferred when oversampling would make your dataset impractically large, or when you want a simpler pipeline. Oversampling (SMOTE, etc.) can generate better decision boundaries in some cases but adds complexity.

Does a higher weight mean a class is more important?

Yes. A weight of 5.0 means the loss for a misclassified sample in that class is multiplied by 5, giving the optimizer five times as strong a gradient signal for that class.

How do I use these weights in scikit-learn?

Pass the dict as class_weight parameter: LogisticRegression(class_weight={0: 1.2, 1: 8.4}). Or use class_weight="balanced" if you want sklearn to compute it automatically from training data.

Can I use these weights in PyTorch or TensorFlow?

Yes. In PyTorch pass the weight tensor to nn.CrossEntropyLoss(weight=...). In TensorFlow/Keras use the class_weight argument in model.fit(). The values from this calculator are directly compatible.

Class Weight Calculator

Name: Class Weight Calculator
Availability: InStock
Author: Nham Vu

Enter class labels and sample counts to compute balanced class weights for imbalanced datasets — scikit-learn compatible output.

Dataset Classes

Enter each class label and the number of samples it contains.

Computed Weights

Enter class data and click Calculate.

Class	Samples	Share	Weight

scikit-learn dict

NumPy / PyTorch tensor

Formula: w_c = n_samples / (n_classes × n_samples_c)

Copied!

Summary

Enter class labels and sample counts to compute balanced class weights for imbalanced datasets — scikit-learn compatible output.

How it works

Enter a class label and its sample count for each class in your dataset.
Add as many classes as needed with the "Add Class" button.
The calculator applies the scikit-learn balanced formula: w = n_total / (n_classes × n_c).
Results appear as a weight dictionary you can paste directly into your training code.
Higher weights indicate rarer classes that the model should prioritize.

Use cases

Binary fraud detection where fraud events are less than 1% of samples.
Medical diagnosis models with rare positive diagnoses.
Multi-class NLP classifiers with unequal label distributions.
Image classification where some categories have fewer photos.
Any sklearn Pipeline using LogisticRegression, SVC, or RandomForestClassifier.
Computing sample_weight arrays for custom PyTorch or TensorFlow loss functions.

Frequently Asked Questions

Last updated: 2026-07-22 · Reviewed by Nham Vu