LOCO

class ordinal_xai.interpretation.loco.LOCO(model, X, y=None, test_size=0.2, random_state=None, use_train_test_split=True)[source]

Bases: BaseInterpretation

Leave-One-Covariate-Out (LOCO) interpretation method for feature importance.

This class implements the LOCO method for analyzing feature importance in ordinal regression models. It measures the impact of each feature by evaluating how model performance changes when the feature is removed from the dataset.

Parameters:

  • model (object) – The trained ordinal regression model. Must implement fit and predict methods.

  • X (pd.DataFrame) – Dataset used for interpretation. Should contain the same features used during model training.

  • y (pd.Series, optional) – Target labels. Required for evaluation but optional for initialization.

  • test_size (float, default=0.2) – Proportion of the dataset to include in the test split when using train-test split evaluation.

  • random_state (int, optional) – Random state for reproducibility of train-test splits.

  • use_train_test_split (bool, default=True) – Whether to use train-test split for evaluation. If False, the entire dataset is used for both training and evaluation.

available_metrics

Dictionary of available evaluation metrics and their corresponding functions.

Type:

dict

test_size

Proportion of data used for testing.

Type:

float

random_state

Random state for reproducibility.

Type:

int

use_train_test_split

Whether to use train-test split.

Type:

bool

X_train, X_test

Training and test datasets when using train-test split.

Type:

pd.DataFrame

y_train, y_test

Training and test labels when using train-test split.

Type:

pd.Series

original_predictions

Predictions from the original model.

Type:

array-like

original_proba_predictions

Probability predictions from the original model if available.

Type:

array-like, optional

original_results

Evaluation results from the original model.

Type:

dict

__init__(model, X, y=None, test_size=0.2, random_state=None, use_train_test_split=True)[source]

Initialize the LOCO interpretation method.

Parameters:

  • model (object) – The trained ordinal regression model

  • X (pd.DataFrame) – Dataset used for interpretation

  • y (pd.Series, optional) – Target labels

  • test_size (float, default=0.2) – Proportion of the dataset to include in the test split

  • random_state (int, optional) – Random state for reproducibility

  • use_train_test_split (bool, default=True) – Whether to use train-test split for evaluation

_perform_train_test_split()[source]

Perform train-test split on the data and evaluate the original model.

This method: 1. Splits the data into training and test sets 2. Creates a copy of the model for training 3. Fits the model on the training data 4. Makes predictions on the test set 5. Evaluates the model using all available metrics

Raises:

ValueError – If the model class is not suitable for refitting

explain(observation_idx=None, feature_subset=None, plot=False, metrics=None, title=True)[source]

Generate LOCO feature importance scores.

This method computes feature importance by measuring how model performance changes when each feature is removed. It can operate in two modes: 1. Global mode (default): Evaluates feature importance across the entire dataset 2. Local mode (when observation_idx is provided): Evaluates feature importance for a specific instance

Parameters:

  • observation_idx (int, optional) – Index of specific instance to analyze (local explanation)

  • feature_subset (list, optional) – List of feature names or indices to analyze

  • plot (bool, default=False) – Whether to create visualizations

  • metrics (list, optional) – List of metrics to use for feature importance calculation

  • title (bool, default=True) – Whether to add a suptitle to the visualization

Returns:

Dictionary containing feature importance scores for each metric

Return type:

dict

_plot_feature_importance(results, metrics=None, title=True)[source]

Create visualizations for feature importance.

This method creates bar plots showing feature importance scores for each metric. The plots are arranged in a grid, with one subplot per metric.

Parameters:

  • results (dict) – Dictionary containing all results from the LOCO analysis

  • metrics (list, optional) – List of metrics to plot (defaults to all in results)

  • title (bool, default=True) – Whether to add a suptitle to the visualization

Notes

  • Features are sorted by importance score

  • Error metrics are shown in red with “Increase” in the title

  • Other metrics are shown in green with “Drop” in the title

  • Feature names are rotated 90 degrees for better readability

  • Score values are displayed in the middle of each bar

_abc_impl = <_abc._abc_data object>