Label Noise Injection Utilities

This module provides functions to simulate mislabeled data, which is useful for evaluating the robustness of classification algorithms under imperfect labeling conditions.

It focuses on binary and categorical (integer-based) classification tasks. The noise is introduced by deliberately altering the target column in a controlled manner, either starting from a clean dataset or extending existing noise.

There are two categories of methods:

• New methods: inject label errors into a clean dataset to reach a given percentage.
• Extended methods: increase the amount of existing label noise until a target percentage is achieved, using a clean reference version.

These tools help researchers and practitioners simulate real-world conditions, such as annotation errors or adversarial mislabeling.

---

wrongLabelsBinaryExtended(original_df, train_df, column, percentage)

Adds label noise to a binary classification target column, extending any existing noise until the desired error rate is reached.

Parameters

• original_df (pd.DataFrame):
  The clean DataFrame used as ground truth reference.

• train_df (pd.DataFrame):
  The dataset that may already contain some label errors.

• column (str):
  The name of the binary target column (values should be 0 or 1).

• percentage (float):
  Final desired proportion (0–1) of incorrect labels.

---

wrongLabelsBinaryNew(train_df, target, percentage)

Injects new binary label noise into a clean dataset by flipping 0 ↔ 1 in a controlled proportion of rows.

Parameters

• train_df (pd.DataFrame):
  The clean dataset to modify.

• target (str):
  The name of the binary target column.

• percentage (float):
  The proportion of values to flip (0–1).

---

wrongLabelsCategoricalNew(train_df, target, percentage)

Introduces label noise into a categorical integer target column by replacing a portion of values with other existing category labels.

Parameters

• train_df (pd.DataFrame):
  The clean dataset.

• target (str):
  The name of the integer-based target column.

• percentage (float):
  Proportion (0–1) of values to replace with incorrect but valid labels.

---

wrongLabelsCategoryExtended(train_df, target, percentage)

Extends existing label noise in a categorical integer target column, replacing additional values with different existing labels until the target error rate is met.

Parameters

• train_df (pd.DataFrame):
  Dataset that may already include some incorrect labels.

• target (str):
  The name of the integer-based target column.

• percentage (float):
  Desired final proportion (0–1) of misclassified labels.