Feature Scaling & Normalization Helper

While often used interchangeably, these terms have specific meanings in data science. **Normalization** typically refers to scaling data to a fixed range like [0, 1] (min-max scaling). **Standardization** (z-score scaling) transforms data to have mean=0 and standard deviation=1. Both are forms of feature scaling, but they use different mathematical transformations. This tool supports both methods so you can compare them directly.

Not all models require feature scaling. **Tree-based models** (Decision Trees, Random Forest, XGBoost, LightGBM) are scale-invariant because they make decisions based on thresholds, not distances. However, **gradient-based models** (Linear/Logistic Regression, Neural Networks, SVM) and **distance-based models** (K-NN, K-Means) typically benefit significantly from scaling. When in doubt, scaling rarely hurts and often helps.

For **classification**, never scale the target—it's categorical. For **regression**, scaling the target is optional but can help with very large or small target values. If you scale the target during training, remember to inverse-transform predictions back to the original scale for interpretation. Neural networks sometimes benefit from scaled targets.

Outliers significantly affect both methods differently. **Min-max scaling** is very sensitive—outliers compress most data into a small range. **Z-score** is more robust but outliers still skew the mean and standard deviation. For data with many outliers, consider **robust scaling** (using median and IQR) or **winsorizing** outliers before scaling. This tool shows you the statistics so you can identify problematic outliers.

Yes! There's no rule that all features must use the same scaler. You might use min-max for features that need bounded output (like neural network inputs) and z-score for features where you want to preserve relative distances. This tool lets you select different methods per feature so you can experiment with mixed approaches.

With **min-max scaling**, new values outside [min, max] will produce scaled values outside [0, 1]—potentially negative or greater than 1. With **z-score**, extreme values will produce larger absolute z-scores than seen in training. Both cases are valid mathematically but may affect model behavior. Some practitioners clip values to training bounds; others let the model handle extrapolation naturally.

This prevents **data leakage**—using information from test/validation data during training. If you compute scaling parameters on the entire dataset, test data statistics "leak" into training, giving overly optimistic performance estimates. Always: (1) split data, (2) fit scaler on training set, (3) transform both train and test using training parameters.

This tool is designed for education. Upload your data and see: (1) computed parameters (mean, std, min, max) for each feature, (2) side-by-side comparison of original and scaled values, (3) visualization of how distributions change, and (4) warnings about potential issues like constant features. Use it to build intuition before applying scaling in real ML pipelines.