Results
Enter data, choose options, then calculate scaling.
Formula used
Scaled dataset
Core method comparison
The first selected feature is compared across four common scaling methods.
Scaling parameters
| Feature | Parameter | Value |
|---|
Step-by-step summary
Visual comparison
Calculator inputs
Inverse transformation
Paste scaled values after calculation. The saved parameters restore original units.
No inverse result yet.
Method guidance
| Method | Best use | Main caution |
|---|---|---|
| Min-Max | Bounded neural network inputs and visual comparisons. | Extreme values can compress most observations. |
| Z-score | Linear models, SVM, PCA, and gradient methods. | Mean and deviation react to large outliers. |
| Robust | Features containing meaningful extreme observations. | Values are not restricted to a fixed range. |
| Max absolute | Sparse data where centring is undesirable. | Large absolute extremes still dominate scaling. |
| Vector norms | Text vectors, embeddings, and directional similarity. | Normalises each vector rather than each feature. |
| Power transforms | Reducing skew before compatible modelling steps. | Domain restrictions and lambda choices matter. |
Algorithms and scaling recommendations
Usually benefits strongly
KNN, SVM, K-Means, PCA, neural networks, logistic regression, and gradient-based models.
Often benefits
Linear regression, regularised models, clustering methods, and distance-based anomaly detection.
Usually less sensitive
Decision trees, random forests, and many boosted tree implementations.
How to use
- Paste data or upload a CSV file.
- Select one dataset or train-and-test mode.
- Choose a scaling method and direction.
- Configure missing values and outlier handling.
- Select columns, precision, and formatting options.
- Calculate, inspect charts, then export results.
Example dataset
| age | income | score | segment |
|---|---|---|---|
| 22 | 32000 | 71 | A |
| 35 | 54000 | 85 | B |
| 48 | 87000 | 93 | A |
| 29 | 41000 | 78 | C |
Frequently asked questions
Why is feature scaling important?
Scaling prevents large numeric ranges from dominating calculations. It improves many distance and gradient methods. It also makes comparisons clearer across different measurement units.
What is the difference between normalisation and standardisation?
Normalisation usually maps values into a chosen range. Standardisation centres values using mean and standard deviation. Their outputs therefore follow different numerical interpretations and constraints.
Which method handles outliers best?
Robust scaling uses the median and percentile spread. These statistics resist extreme observations better than means. Outlier treatment should still respect the dataset's meaning.
Should tree models use scaled features?
Tree splits generally depend on ordering rather than distance. Scaling often changes little for tree predictions. It may still help shared preprocessing pipelines and diagnostics.
How does train-test mode prevent leakage?
The calculator fits parameters only from training rows. Testing rows use those unchanged training parameters. This prevents testing information influencing the fitted transformation.
Can categorical columns remain unchanged?
Yes, categorical columns can be preserved automatically. Identifier columns can also remain untouched. Only selected numeric features receive the chosen scaling transformation.
What happens with constant features?
A constant feature has zero range or deviation. Several methods cannot scale it meaningfully. The calculator returns zero and reports the condition clearly.
Can scaled values be restored?
Invertible methods can restore values using stored parameters. Log and power methods also support inverse calculations. Row-wise transformations require matching row parameters for accurate restoration.
Which transformations require positive values?
Logarithms, square roots, reciprocals, and Box-Cox have restrictions. Log1p accepts values greater than negative one. Yeo-Johnson handles zero and negative values more safely.