Quantify class-wise signal strength for every numeric predictor. Review means, sums, p-values, and effect size. Build cleaner datasets with confident evidence driven feature ranking.
Paste one numeric feature across target classes. Separate values with commas, spaces, or new lines.
Use this sample to test the calculator. It represents one numeric feature measured across three target classes.
| Class | Observations |
|---|---|
| Spam | 7.2, 6.8, 7.5, 6.9, 7.1 |
| Promotions | 5.8, 5.9, 6.1, 6.0, 5.7 |
| Personal | 8.2, 8.0, 7.9, 8.4, 8.1 |
Grand Mean: \( \bar{x} = \frac{\sum x}{N} \)
Between Class Sum of Squares: \( SSB = \sum n_i(\bar{x_i} - \bar{x})^2 \)
Within Class Sum of Squares: \( SSW = \sum \sum (x_{ij} - \bar{x_i})^2 \)
Degrees of Freedom: \( df_{between} = k - 1 \), \( df_{within} = N - k \)
Mean Squares: \( MSB = \frac{SSB}{df_{between}} \), \( MSW = \frac{SSW}{df_{within}} \)
ANOVA F Score: \( F = \frac{MSB}{MSW} \)
Effect Size: \( \eta^2 = \frac{SSB}{SST} \)
In machine learning feature selection, a larger F score usually indicates stronger discrimination between classes for that numeric feature.
It measures how strongly a numeric feature separates target classes. A larger F score means between-class differences are large relative to within-class variation.
Use it during filter-based feature selection for classification tasks, especially when you want quick evidence about whether class means differ for one numeric predictor.
Yes. Run one feature at a time, record each F score, and compare the results. Higher scores often point to more informative predictors.
The p value estimates how likely your observed class differences would appear if all class means were actually equal. Smaller values suggest stronger evidence.
Critical F gives a threshold based on your significance level and degrees of freedom. When the computed F exceeds that threshold, class separation is statistically significant.
Eta squared is an effect size. It shows how much of the total feature variance is explained by class membership, helping you judge practical importance.
Yes. Observations should be independent, groups should be approximately normal, and variances should be reasonably similar. Severe violations can weaken conclusions.
The calculator can still process data if overall degrees of freedom remain valid, but very small groups reduce reliability. More observations usually produce better estimates.
Important Note: All the Calculators listed in this site are for educational purpose only and we do not guarentee the accuracy of results. Please do consult with other sources as well.