interpretable 0.0.1

interpretable:

interpretable

Interpretable machine learning toolbox.
Examples · Explore the API

Examples

🕸️ Graph Neural Network based auto-encoder
🛠️ Preprocessing.
📦 Learning.
🧪 Evaluation.
🔢 Interpretation.

Installation

pip install interpretable              # with basic dependencies  

With additional dependencies as required:

pip install interpretable[ml]          # for machine learning applications using scikit-learn
pip install interpretable[dl]          # for deep learning based applications using pytorch 
pip install interpretable[gnn]         # for graph neural network based applications using pytorch geometric 
pip install interpretable[dev]         # for local testing

How to cite?

Please cite it using the metadata given in this file. For more information about citation, please see 'Cite this repository' section on the github page of the repository.

Future directions, for which contributions are welcome:

• Support for classification models other than RFC and GBC.
• Support for regression models.
• More examples of GNNs.

Similar projects:

• https://github.com/EpistasisLab/tpot
• https://github.com/oegedijk/explainerdashboard

API

module interpretable.gnn.layers

---

function get_channels

get_channels(start, end, scale, kind)

---

function get_layers

get_layers(
    model_name,
    num_node_features,
    hidden_channels,
    kind,
    scale,
    **kws_model
)

Get the layers for encoding or decoding.

---

function get_coder

get_coder(
    model_name,
    num_node_features,
    hidden_channels,
    kind,
    scale,
    **kws_model
)

Get a stack of layers for encoding or decoding

module interpretable.gnn

module interpretable.ml.classify

For classification.

---

function get_grid_search

get_grid_search(
    modeln: str,
    X: <built-in function array>,
    y: <built-in function array>,
    param_grid: dict = {},
    cv: int = 5,
    n_jobs: int = 6,
    random_state: int = None,
    scoring: str = 'balanced_accuracy',
    **kws
) → object

Grid search.

Args:

• modeln (str): name of the model.
• X (np.array): X matrix.
• y (np.array): y vector.
• param_grid (dict, optional): parameter grid. Defaults to {}.
• cv (int, optional): cross-validations. Defaults to 5.
• n_jobs (int, optional): number of cores. Defaults to 6.
• random_state (int, optional): random state. Defaults to None.
• scoring (str, optional): scoring system. Defaults to 'balanced_accuracy'.

Keyword arguments:

• kws: parameters provided to the GridSearchCV function.

Returns:

• object: grid_search.

References:

• 1. https: //scikit-learn.org/stable/modules/generated/sklearn.model_selection.GridSearchCV.html
• 2. https: //scikit-learn.org/stable/modules/model_evaluation.html

---

function get_estimatorn2grid_search

get_estimatorn2grid_search(
    estimatorn2param_grid: dict,
    X: DataFrame,
    y: Series,
    **kws
) → dict

Estimator-wise grid search.

Args:

• estimatorn2param_grid (dict): estimator name to the grid search map.
• X (pd.DataFrame): X matrix.
• y (pd.Series): y vector.

Returns:

• dict: output.

---

function get_test_scores

get_test_scores(d1: dict) → DataFrame

Test scores.

Args:

• d1 (dict): dictionary with objects.

Returns:

• pd.DataFrame: output.

TODOs: Get best param index.

---

function run_grid_search

run_grid_search(
    df,
    cols_x,
    coly,
    colindex,
    n_estimators: int = None,
    qcut: float = None,
    evaluations: list = ['prediction', 'feature importances', 'partial dependence'],
    estimatorn2param_grid: dict = None,
    output_dir_path: str = None,
    test: bool = False,
    **kws
) → dict

Run grid search.

Args:

• n_estimators (int): number of estimators.
• qcut (float, optional): quantile cut-off. Defaults to None.
• evaluations (list, optional): evaluations types. Defaults to ['prediction','feature importances', 'partial dependence', ].
• estimatorn2param_grid (dict, optional): estimator to the parameter grid map. Defaults to None.
• output_dir_path (str, optional): output_dir_pathut path. Defaults to None.
• test (bool, optional): test mode. Defaults to False.

Keyword arguments:

• kws: parameters provided to get_estimatorn2grid_search.

Returns:

• dict: estimator to grid search map.

module interpretable.ml.evaluate

---

function get_probability

get_probability(
    estimatorn2grid_search: dict,
    X: <built-in function array>,
    y: <built-in function array>,
    colindex: str,
    coff: float = 0.5,
    test: bool = False
) → DataFrame

Classification probability.

Args:

• estimatorn2grid_search (dict): estimator to the grid search map.
• X (np.array): X matrix.
• y (np.array): y vector.
• colindex (str): index column.
• coff (float, optional): cut-off. Defaults to 0.5.
• test (bool, optional): test mode. Defaults to False.

Returns:

• pd.DataFrame: output.

---

function get_auc_cv

get_auc_cv(estimator, X, y, cv=5, test=False, fitted=False, random_state=None)

TODO: just predict_probs as inputs TODO: resolve duplication of stat.binary.auc TODO: add more metrics in ds1 in addition to auc

---

function get_roc_auc

get_roc_auc(true, test, outmore=False)

module interpretable.ml.interpret

---

function get_feature_predictive_power

get_feature_predictive_power(
    d0: dict,
    df01: DataFrame,
    n_splits: int = 5,
    n_repeats: int = 10,
    random_state: int = None,
    plot: bool = False,
    drop_na: bool = False,
    **kws
) → DataFrame

get_feature_predictive_power summary

Notes:

x-values should be scale and sign agnostic.

Args:

• d0 (dict): input dictionary.
• df01 (pd.DataFrame): input data,
• n_splits (int, optional): number of splits. Defaults to 5.
• n_repeats (int, optional): number of repeats. Defaults to 10.
• random_state (int, optional): random state. Defaults to None.
• plot (bool, optional): plot. Defaults to False.
• drop_na (bool, optional): drop missing values. Defaults to False.

Returns:

• pd.DataFrame: output data.

---

function get_feature_importances

get_feature_importances(
    estimatorn2grid_search: dict,
    X: DataFrame,
    y: Series,
    scoring: str = 'roc_auc',
    n_repeats: int = 20,
    n_jobs: int = 6,
    random_state: int = None,
    plot: bool = False,
    test: bool = False,
    **kws
) → DataFrame

Feature importances.

Args:

• estimatorn2grid_search (dict): map between estimator name and grid search object.
• X (pd.DataFrame): X matrix.
• y (pd.Series): y vector.
• scoring (str, optional): scoring type. Defaults to 'roc_auc'.
• n_repeats (int, optional): number of repeats. Defaults to 20.
• n_jobs (int, optional): number of cores. Defaults to 6.
• random_state (int, optional): random state. Defaults to None.
• plot (bool, optional): plot. Defaults to False.
• test (bool, optional): test mode. Defaults to False.

Returns:

• pd.DataFrame: output data.

---

function get_partial_dependence

get_partial_dependence(
    estimatorn2grid_search: dict,
    X: DataFrame,
    y: Series,
    test: bool = False
) → DataFrame

Partial dependence.

Args:

• estimatorn2grid_search (dict): map between estimator name and grid search object.
• X (pd.DataFrame): X matrix.
• y (pd.Series): y vector.

Returns:

• pd.DataFrame: output data.

---

function interpret

interpret(
    input_dir_path,
    output_dir_path,
    keys=['predictive power', 'feature importances', 'partial dependence', 'feature contributions'],
    random_state=None,
    plot=False,
    test=False
)

---

function agg_predictive_power

agg_predictive_power(df)

---

function agg_feature_importances

agg_feature_importances(df)

---

function agg_feature_contributions

agg_feature_contributions(df4)

---

function agg_feature_interpretations

agg_feature_interpretations(interprets: dict)

module interpretable.ml.io

---

function read_models

read_models(
    output_dir_path,
    keys=None,
    filenames={'inputs': 'input.json', 'data': 'input.pqt', 'estimators': 'estimatorn2grid_search.pickle', 'predictions': 'prediction.pqt'}
)

module interpretable.ml

module interpretable.ml.pre

---

function get_Xy

get_Xy(df01, columns, y_kind)

Get the columns for a kind of model

---

function get_Xy_for_classification

get_Xy_for_classification(
    df1: DataFrame,
    coly: str,
    qcut: float = None,
    drop_xs_low_complexity: bool = False,
    min_nunique: int = 5,
    max_inflation: float = 0.5,
    **kws
) → dict

Get X matrix and y vector.

Args:

• df1 (pd.DataFrame): input data, should be indexed.
• coly (str): column with y values, bool if qcut is None else float/int
• qcut (float, optional): quantile cut-off. Defaults to None.
• drop_xs_low_complexity (bool, optional): to drop columns with <5 unique values. Defaults to False.
• min_nunique (int, optional): minimum unique values in the column. Defaults to 5.
• max_inflation (float, optional): maximum inflation. Defaults to 0.5.

Keyword arguments:

• kws: parameters provided to drop_low_complexity.

Returns:

• dict: output.

module interpretable.viz.annot

---

function annot_confusion_matrix

annot_confusion_matrix(df_: DataFrame, ax: Axes = None, off: float = 0.5) → Axes

Annotate a confusion matrix.

Args:

• df_ (pd.DataFrame): input data.
• ax (plt.Axes, optional): plt.Axes object. Defaults to None.
• off (float, optional): offset. Defaults to 0.5.

Returns:

• plt.Axes: plt.Axes object.

module interpretable.viz.gnn

---

function lines_metricsby_epochs

lines_metricsby_epochs(data, figsize=[3, 3])

Args:

• data: table containing the epoch and other metrics.

module interpretable.viz

module interpretable.viz.ml

---

function plot_metrics

plot_metrics(data, inputs, estimators, plot: bool = False) → DataFrame

Plot performance metrics.

Args:

• plot (bool, optional): make plots. Defaults to False.

Returns:

• pd.DataFrame: output data.

---

function plot_feature_predictive_power

plot_feature_predictive_power(
    df3: DataFrame,
    ax: Axes = None,
    figsize: list = [3, 3],
    **kws
) → Axes

Plot feature-wise predictive power.

Args:

• df3 (pd.DataFrame): input data.
• ax (plt.Axes, optional): axes object. Defaults to None.
• figsize (list, optional): figure size. Defaults to [3,3].

Returns:

• plt.Axes: output.

---

function plot_feature_ranks

plot_feature_ranks(df2: DataFrame)

---

function plot_feature_contributions

plot_feature_contributions(data, kws_plot, vmax=0.2, vmin=-0.2, figsize=[4, 4])