target_leakage_data_check

Data check that checks if any of the features are highly correlated with the target by using mutual information or Pearson correlation.

Module Contents

Classes Summary

TargetLeakageDataCheck

Check if any of the features are highly correlated with the target by using mutual information, Pearson correlation, and other correlation metrics.

Contents

class evalml.data_checks.target_leakage_data_check.TargetLeakageDataCheck(pct_corr_threshold=0.95, method='mutual_info')

Check if any of the features are highly correlated with the target by using mutual information, Pearson correlation, and other correlation metrics.

If method=’mutual_info’, this data check uses mutual information and supports all target and feature types. Other correlation metrics only support binary with numeric and boolean dtypes. This method will return a value in [-1, 1] if other correlation metrics are selected and will returns a value in [0, 1] if mutual information is selected. Correlation metrics available can be found in Woodwork’s dependence_dict method.

Parameters

• pct_corr_threshold (float) – The correlation threshold to be considered leakage. Defaults to 0.95.

• method (string) – The method to determine correlation. Use ‘max’ for the maximum correlation, or for specific correlation metrics, use their name (ie ‘mutual_info’ for mutual information, ‘pearson’ for Pearson correlation, etc). possible methods can be found in Woodwork’s config, under correlation_metrics. Excludes ‘all’. Defaults to ‘mutual_info’.

Methods

name	Return a name describing the data check.
validate	Check if any of the features are highly correlated with the target by using mutual information, Pearson correlation, and/or Spearman correlation.

name(cls)

Return a name describing the data check.

validate(self, X, y)

Check if any of the features are highly correlated with the target by using mutual information, Pearson correlation, and/or Spearman correlation.

If method=’mutual_info’ or ‘method=’max’, supports all target and feature types. Other correlation metrics only support binary with numeric and boolean dtypes. This method will return a value in [-1, 1] if other correlation metrics are selected and will returns a value in [0, 1] if mutual information is selected.

Parameters

• X (pd.DataFrame, np.ndarray) – The input features to check.

• y (pd.Series, np.ndarray) – The target data.

Returns

dict with a DataCheckWarning if target leakage is detected.

Return type

dict (DataCheckWarning)

Examples

>>> import pandas as pd

Any columns that are strongly correlated with the target will raise a warning. This could be indicative of data leakage.

>>> X = pd.DataFrame({
...    "leak": [10, 42, 31, 51, 61] * 15,
...    "x": [42, 54, 12, 64, 12] * 15,
...    "y": [13, 5, 13, 74, 24] * 15,
... })
>>> y = pd.Series([10, 42, 31, 51, 40] * 15)
...
>>> target_leakage_check = TargetLeakageDataCheck(pct_corr_threshold=0.95)
>>> assert target_leakage_check.validate(X, y) == [
...     {
...         "message": "Column 'leak' is 95.0% or more correlated with the target",
...         "data_check_name": "TargetLeakageDataCheck",
...         "level": "warning",
...         "code": "TARGET_LEAKAGE",
...         "details": {"columns": ["leak"], "rows": None},
...         "action_options": [
...             {
...                 "code": "DROP_COL",
...                 "data_check_name": "TargetLeakageDataCheck",
...                 "parameters": {},
...                 "metadata": {"columns": ["leak"], "rows": None}
...             }
...         ]
...     }
... ]

The default method can be changed to pearson from mutual_info.

>>> X["x"] = y / 2
>>> target_leakage_check = TargetLeakageDataCheck(pct_corr_threshold=0.8, method="pearson")
>>> assert target_leakage_check.validate(X, y) == [
...     {
...         "message": "Columns 'leak', 'x' are 80.0% or more correlated with the target",
...         "data_check_name": "TargetLeakageDataCheck",
...         "level": "warning",
...         "details": {"columns": ["leak", "x"], "rows": None},
...         "code": "TARGET_LEAKAGE",
...         "action_options": [
...             {
...                 "code": "DROP_COL",
...                  "data_check_name": "TargetLeakageDataCheck",
...                  "parameters": {},
...                  "metadata": {"columns": ["leak", "x"], "rows": None}
...             }
...         ]
...     }
... ]