RobustScaler

RobustScaler is a data preprocessing transformer in scikit-learn designed to scale features in a way that is resistant to outliers. It is commonly used in machine learning pipelines to normalize data without being unduly influenced by extreme values, making it preferable for datasets with skewed distributions or outliers.

The transformation works by centering and scaling each feature using robust statistics. Specifically, it subtracts the feature-wise median and then divides by the interquartile range (IQR) defined by a specified quantile range, typically the 25th to 75th percentiles. The resulting data have reduced impact from outliers because the median and IQR are less sensitive to extreme values than the mean and standard deviation. The transformation is applied independently to each feature.

Key parameters include:

- with_center: whether to center the data by subtracting the median (default True).

- with_scaling: whether to scale by the IQR (default True).

- quantile_range: the lower and upper quantiles used to compute the IQR (default (25.0, 75.0)).

- copy: whether to return a copy of the input data (default True).

RobustScaler is suitable when features contain outliers or heavy-tailed distributions, providing a more robust alternative to StandardScaler, which uses mean and standard deviation. It can be integrated into pipelines alongside estimators and is applicable to dense and, in many cases, sparse input. In practice, it is often paired with tree-based or linear models to improve performance when data robustness is essential.