Calculated Results
Results update after you calculate or change valid settings.
| Class | Count | Weighted count | Percentage | Cumulative | Rank | Balanced weight | Status |
|---|
Interpretation
Recommendations
- Calculate the distribution to receive recommendations.
Result Summary
No calculated summary yet.
Formula Used
Class percentage: pᵢ = nᵢ ÷ N × 100.
Imbalance ratio: IR = max(nᵢ) ÷ min(nᵢ).
Entropy: H = −Σ pᵢ ln(pᵢ).
Normalized entropy: H ÷ ln(K).
Gini impurity: 1 − Σ pᵢ².
Balanced weight: wᵢ = N ÷ (K × nᵢ).
Here, N is total samples. K is total classes. The value nᵢ is one class count.
How to Use
- Add class names and counts.
- Enter split or prediction counts when available.
- Select sorting, filters, and chart options.
- Choose resampling and class-weight methods.
- Select Calculate distribution.
- Review imbalance, drift, and rare-class warnings.
- Export the chart or calculated report.
Example Data Table
| Class | Overall | Train | Validation | Test | Predicted |
|---|---|---|---|---|---|
| Cat | 520 | 360 | 80 | 80 | 490 |
| Dog | 410 | 285 | 65 | 60 | 430 |
| Bird | 145 | 100 | 20 | 25 | 155 |
| Horse | 75 | 50 | 10 | 15 | 75 |
Frequently Asked Questions
What is a class distribution?
It shows samples assigned to each target class. It reveals whether categories have similar representation. Large differences may influence model learning.
Why does class imbalance matter?
Models may favor large classes during training. Accuracy can appear strong despite weak minority recall. Balanced metrics give a clearer evaluation.
What is a majority-class baseline?
It predicts the largest class every time. This provides a simple accuracy benchmark. Better models should exceed this baseline meaningfully.
What imbalance ratio is acceptable?
No single threshold fits every problem. Costs and sample sizes also matter. Review recall and precision for each class.
Should I always oversample minorities?
No method works for every dataset. Oversampling may copy noise or outliers. Validate choices using unseen data.
When should class weights be used?
Weights help many loss-based classification models. They increase minority errors during optimization. Compare weighted and unweighted validation results.
What does normalized entropy show?
It measures how evenly classes are distributed. Values near one indicate stronger balance. Values near zero indicate concentration.
How is split drift detected?
The calculator compares class percentages across splits. Large percentage-point differences trigger warnings. Missing classes receive separate warnings.
Can this replace model evaluation?
No, distribution analysis is only diagnostic. Use confusion matrices and per-class metrics. Evaluate performance on representative unseen data.