Outliers

outliers.py injects artificial outliers into datasets. It is designed for testing the resilience of machine learning models against extreme values and abnormal input data.

Outlier injection uses a 3-sigma rule for continuous and discrete numeric data, domain expansion for categorical integers, and a fixed string token ("puck was here") for text columns.

Function Signature

from pucktrick.outliers import outlier

error_code, modified_df = outlier(df, strategy)
# or, for mode="extended" / mode="composed":
error_code, modified_df = outlier(df, strategy, original_df=clean_df)

Strategy Parameters

No module-specific parameters are required inside perturbate_data. The injector automatically detects column type and applies the appropriate outlier logic:

Column type	Outlier method
Continuous numeric	Values outside μ ± 3σ
Discrete numeric	Integer values outside the 3-sigma integer range
Categorical integer	Integers outside the existing min/max range
Categorical string	Replaced with "puck was here"

Example

from pucktrick.outliers import outlier

strategy = {
    "affected_features": ["temperature", "pressure"],
    "selection_criteria": "all",
    "percentage": 0.10,
    "mode": "new",
    "perturbate_data": {"sampling": "random", "distribution": "random"}
}

err, df_corrupted = outlier(df, strategy)

Modes

Mode	Behaviour
new	Injects outliers into a clean dataset up to the specified percentage.
extended	Adds additional outliers to columns that may already contain some, reaching the cumulative percentage target. Requires original_df.
composed	Injects outliers only into rows already modified by a previous operator. Requires original_df.