Skip to main content
Version: 1.0.0

Noisy

This module provides a set of utility functions for injecting synthetic noise into a dataset.
These functions are useful for testing the robustness of machine learning models under various types of data corruption, such as:

  • Categorical perturbations (existing or fake values)
  • Numerical noise (discrete, continuous, or binary flips)

Each noise type includes two variants:

  • New methods: Inject noise into a clean column from scratch to reach the specified percentage.
  • Extended methods: Add additional noise to an already partially noisy column, up to the desired percentage.

The module supports:

  • Continuous numerical features (noiseContinueNew, noiseContinueExtended)
  • Discrete numerical features (noiseDiscreteNew, noiseDiscreteExtended)
  • Binary features (noiseBinaryNew, noiseBinaryExtended)
  • Categorical (string and integer) features, with:
    • Substitution by existing values (noiseCategoricalStringNewExistingValues, etc.)
    • Substitution by fake/random values (noiseCategoricalStringNewFakeValues, etc.)

These functions assume access to both the clean dataset (original_df) and the dataset to be modified (train_df) when using Extended variants.

noiseCategoricalStringNewExistingValues(train_df, column, percentage)​

Replaces a specified percentage of string values in the column with different values randomly selected from the set of existing unique values.

Parameters​

  • train_df (pd.DataFrame):
    The DataFrame containing the original data.

  • column (str):
    The name of the string column to introduce noise in.

  • percentage (float):
    The proportion of entries to modify (range: 0 to 1).

noiseCategoricalStringExtendedExistingValues(original_df, train_df, column, percentage)​

Adds more noise to a string column already containing noisy data, replacing values with other existing ones until the desired percentage is reached.

Parameters​

  • original_df (pd.DataFrame):
    The original (clean) DataFrame before any noise is added.

  • train_df (pd.DataFrame):
    The DataFrame that may already include noisy values in the column.

  • column (str):
    The name of the string column to extend noise in.

  • percentage (float):
    The final target proportion of noisy entries (0–1).

noiseCategoricalStringNewFakeValues(train_df, column, percentage)​

Replaces a specified percentage of string values in the column with randomly generated fake strings (e.g., random sequences of letters).

Parameters​

  • train_df (pd.DataFrame):
    The DataFrame containing the original data.

  • column (str):
    The name of the string column to modify.

  • percentage (float):
    The fraction of rows to replace with fake values (0–1).

noiseCategoricalStringExtendedFakeValues(original_df, train_df, column, percentage)​

Extends fake noise in a string column by injecting additional randomly generated values to reach the target noise percentage.

Parameters​

  • original_df (pd.DataFrame):
    The clean baseline DataFrame.

  • train_df (pd.DataFrame):
    The DataFrame with possible existing fake values.

  • column (str):
    The string column to extend with fake noise.

  • percentage (float):
    The final noise target percentage (0–1).

noiseCategoricalIntNewExistingValues(train_df, column, percentage)​

Replaces a percentage of integer categorical values in the column with other existing values from the same column.

Parameters​

  • train_df (pd.DataFrame):
    Input DataFrame.

  • column (str):
    The integer categorical column to inject noise into.

  • percentage (float):
    Proportion of rows to modify (0–1).

noiseCategoricalIntExtendedExistingValues(original_df, train_df, column, percentage)​

Extends the proportion of noise in an integer categorical column by replacing additional values with other existing ones from the original distribution.

Parameters​

  • original_df (pd.DataFrame):
    The clean original DataFrame.

  • train_df (pd.DataFrame):
    The DataFrame to be extended with noise.

  • column (str):
    The integer categorical column to target.

  • percentage (float):
    Final target proportion of noisy entries (0–1).

noiseDiscreteNew(train_df, column, percentage)​

Replaces a percentage of values in a discrete numerical column with new randomly generated integer values within the column’s original range.

Parameters​

  • train_df (pd.DataFrame):
    Input DataFrame.

  • column (str):
    The discrete integer column to modify.

  • percentage (float):
    Fraction of rows to be replaced (0–1).

noiseDiscreteExtended(original_df, train_df, column, percentage)​

Extends noise in a discrete numerical column by adding new randomly generated integer values to reach the desired level of distortion.

Parameters​

  • original_df (pd.DataFrame):
    The original clean dataset.

  • train_df (pd.DataFrame):
    Dataset which may already include noisy entries.

  • column (str):
    Name of the integer/discrete column to target.

  • percentage (float):
    Desired final noise percentage (0–1).

noiseBinaryNew(train_df, column, percentage)​

Flips the value (0 ↔ 1) of a specified percentage of entries in a binary column.

Parameters​

  • train_df (pd.DataFrame):
    Input DataFrame.

  • column (str):
    The name of the binary (0/1) column to apply flips.

  • percentage (float):
    Proportion of rows to flip (0–1).

noiseBinaryExtended(original_df, train_df, column, percentage)​

Adds additional binary noise (0 ↔ 1 flips) to an already partially noisy binary column to meet a specified overall noise level.

Parameters​

  • original_df (pd.DataFrame):
    Clean baseline DataFrame.

  • train_df (pd.DataFrame):
    Dataset with possible pre-existing flips.

  • column (str):
    Name of the binary column to modify.

  • percentage (float):
    Target noise proportion (0–1).

noiseContinueNew(train_df, column, percentage)​

Injects continuous noise by replacing a percentage of numeric values with new random values uniformly sampled from the column’s original range.

Parameters​

  • train_df (pd.DataFrame):
    Input DataFrame.

  • column (str):
    The numeric (continuous) column to alter.

  • percentage (float):
    Fraction of values to replace with noise (0–1).

noiseContinueExtended(original_df, train_df, column, percentage)​

Extends continuous noise in a numeric column by injecting additional values until the target noise percentage is reached.

Parameters​

  • original_df (pd.DataFrame):
    The original clean dataset.

  • train_df (pd.DataFrame):
    Dataset possibly containing some noise already.

  • column (str):
    The numeric column to apply noise to.

  • percentage (float):
    Final desired noise proportion (0–1).