def visualize(shap_values, feature_names=None, data=None, out_names=None):
    """ Visualize the given SHAP values with an additive force layout. """

    if type(shap_values) != np.ndarray:
        return iml.visualize(shap_values)

    if len(shap_values.shape) == 1:
        shap_values = np.reshape(shap_values, (1,len(shap_values)))

    if out_names is None:
        out_names = ["output value"]

    if shap_values.shape[0] == 1:
        if feature_names is None:
            feature_names = ["" for i in range(shap_values.shape[1]-1)]
        if data is None:
            data = ["" for i in range(len(feature_names))]
        if type(data) == np.ndarray:
            data = data.flatten()

        instance = Instance(np.zeros((1,len(feature_names))), data)
        e = AdditiveExplanation(
            shap_values[0,-1],
            np.sum(shap_values[0,:]),
            shap_values[0,:-1],
            None,
            instance,
            IdentityLink(),
            Model(None, out_names),
            DenseData(np.zeros((1,len(feature_names))), list(feature_names))
        )
        return e

    else:
        exps = []
        for i in range(shap_values.shape[0]):
            if feature_names is None:
                feature_names = ["" for i in range(shap_values.shape[1]-1)]
            if data is None:
                display_data = ["" for i in range(len(feature_names))]
            else:
                display_data = data[i,:]

            instance = Instance(np.ones((1,len(feature_names))), display_data)
            e = AdditiveExplanation(
                shap_values[i,-1],
                np.sum(shap_values[i,:]),
                shap_values[i,:-1],
                None,
                instance,
                IdentityLink(),
                Model(None, out_names),
                DenseData(np.ones((1,len(feature_names))), list(feature_names))
            )
            exps.append(e)
        return exps
Exemple #2
0
def explain_instances(model, data, feature_names, out_names):
    if out_names is None:
        out_names = ["model output"]
    if feature_names is None:
        feature_names = [(i + 1) + "" for i in range(data.shape[1])]

    if type(model) == xgboost.core.Booster:
        exps = []
        contribs = model.predict(xgboost.DMatrix(data), pred_contribs=True)
        for i in range(data.shape[0]):
            instance = Instance(data[i:i + 1, :], data[i, :])
            e = AdditiveExplanation(
                contribs[i, -1], np.sum(contribs[i, :]), contribs[i, :-1],
                None, instance, IdentityLink(), Model(None, out_names),
                DenseData(np.zeros((1, data.shape[1])), list(feature_names)))
            exps.append(e)
        return exps
Exemple #3
0
def explain_instance(model, data, feature_names, out_names):
    if out_names is None:
        out_names = ["model output"]
    if feature_names is None:
        feature_names = [(i + 1) + "" for i in range(data.shape[1])]

    if type(model) == xgboost.core.Booster:
        contribs = model.predict(xgboost.DMatrix(data), pred_contribs=True)
    elif type(model) == lightgbm.basic.Booster:
        contribs = model.predict(data, pred_contrib=True)
    else:
        return None

    instance = Instance(data[0:1, :], data[0, :])
    e = AdditiveExplanation(
        contribs[0, -1], np.sum(contribs[0, :]), contribs[0, :-1], None,
        instance, IdentityLink(), Model(None, out_names),
        DenseData(np.zeros((1, data.shape[1])), list(feature_names)))
    return e
Exemple #4
0
def force_plot(shap_values, features=None, feature_names=None, out_names=None, link="identity",
               plot_cmap="RdBu"):
    """ Visualize the given SHAP values with an additive force layout. """

    assert not type(shap_values) == list, "The shap_values arg looks looks multi output, try shap_values[i]."

    link = iml.links.convert_to_link(link)

    if type(shap_values) != np.ndarray:
        return iml.visualize(shap_values)

    # convert from a DataFrame or other types
    if str(type(features)) == "<class 'pandas.core.frame.DataFrame'>":
        if feature_names is None:
            feature_names = list(features.columns)
        features = features.as_matrix()
    elif str(type(features)) == "<class 'pandas.core.series.Series'>":
        if feature_names is None:
            feature_names = list(features.index)
        features = features.as_matrix()
    elif str(type(features)) == "list":
        if feature_names is None:
            feature_names = features
        features = None
    elif features is not None and len(features.shape) == 1 and feature_names is None:
        feature_names = features
        features = None

    if len(shap_values.shape) == 1:
        shap_values = np.reshape(shap_values, (1, len(shap_values)))

    if out_names is None:
        out_names = ["output value"]

    if shap_values.shape[0] == 1:
        if feature_names is None:
            feature_names = ["Feature " + str(i) for i in range(shap_values.shape[1] - 1)]
        if features is None:
            features = ["" for _ in range(len(feature_names))]
        if type(features) == np.ndarray:
            features = features.flatten()

        instance = Instance(np.zeros((1, len(feature_names))), features)
        e = AdditiveExplanation(
            shap_values[0, -1],
            np.sum(shap_values[0, :]),
            shap_values[0, :-1],
            None,
            instance,
            link,
            Model(None, out_names),
            DenseData(np.zeros((1, len(feature_names))), list(feature_names))
        )
        return iml.visualize(e, plot_cmap)

    else:
        if shap_values.shape[0] > 3000:
            warnings.warn("shap.force_plot is slow many thousands of rows, try subsampling your data.")

        exps = []
        for i in range(shap_values.shape[0]):
            if feature_names is None:
                feature_names = ["Feature " + str(i) for i in range(shap_values.shape[1] - 1)]
            if features is None:
                display_features = ["" for i in range(len(feature_names))]
            else:
                display_features = features[i, :]

            instance = Instance(np.ones((1, len(feature_names))), display_features)
            e = AdditiveExplanation(
                shap_values[i, -1],
                np.sum(shap_values[i, :]),
                shap_values[i, :-1],
                None,
                instance,
                link,
                Model(None, out_names),
                DenseData(np.ones((1, len(feature_names))), list(feature_names))
            )
            exps.append(e)
        return iml.visualize(exps, plot_cmap=plot_cmap)
Exemple #5
0
def force_plot(base_value,
               shap_values,
               features=None,
               feature_names=None,
               out_names=None,
               link="identity",
               plot_cmap="RdBu"):
    """ Visualize the given SHAP values with an additive force layout. """

    # auto unwrap the base_value
    if type(base_value) == np.ndarray and len(base_value) == 1:
        base_value = base_value[0]

    if (type(base_value) == np.ndarray or type(base_value) == list):
        if type(shap_values) != list or len(shap_values) != len(base_value):
            raise Exception("In v0.20 force_plot now requires the base value as the first parameter! " \
                            "Try shap.force_plot(explainer.expected_value, shap_values) or " \
                            "for multi-output models try " \
                            "shap.force_plot(explainer.expected_value[0], shap_values[0]).")

    assert not type(
        shap_values
    ) == list, "The shap_values arg looks looks multi output, try shap_values[i]."

    link = iml.links.convert_to_link(link)

    if type(shap_values) != np.ndarray:
        return iml.visualize(shap_values)

    # convert from a DataFrame or other types
    if str(type(features)) == "<class 'pandas.core.frame.DataFrame'>":
        if feature_names is None:
            feature_names = list(features.columns)
        features = features.values
    elif str(type(features)) == "<class 'pandas.core.series.Series'>":
        if feature_names is None:
            feature_names = list(features.index)
        features = features.values
    elif isinstance(features, list):
        if feature_names is None:
            feature_names = features
        features = None
    elif features is not None and len(
            features.shape) == 1 and feature_names is None:
        feature_names = features
        features = None

    if len(shap_values.shape) == 1:
        shap_values = np.reshape(shap_values, (1, len(shap_values)))

    if out_names is None:
        out_names = ["output value"]

    if shap_values.shape[0] == 1:
        if feature_names is None:
            feature_names = [
                labels['FEATURE'] % str(i) for i in range(shap_values.shape[1])
            ]
        if features is None:
            features = ["" for _ in range(len(feature_names))]
        if type(features) == np.ndarray:
            features = features.flatten()

        # check that the shape of the shap_values and features match
        if len(features) != shap_values.shape[1]:
            msg = "Length of features is not equal to the length of shap_values!"
            if len(features) == shap_values.shape[1] - 1:
                msg += " You might be using an old format shap_values array with the base value " \
                       "as the last column. In this case just pass the array without the last column."
            raise Exception(msg)

        instance = Instance(np.zeros((1, len(feature_names))), features)
        e = AdditiveExplanation(
            base_value,
            np.sum(shap_values[0, :]) + base_value, shap_values[0, :], None,
            instance, link, Model(None, out_names),
            DenseData(np.zeros((1, len(feature_names))), list(feature_names)))
        return iml.visualize(e, plot_cmap)

    else:
        if shap_values.shape[0] > 3000:
            warnings.warn(
                "shap.force_plot is slow many thousands of rows, try subsampling your data."
            )

        exps = []
        for i in range(shap_values.shape[0]):
            if feature_names is None:
                feature_names = [
                    labels['FEATURE'] % str(i)
                    for i in range(shap_values.shape[1])
                ]
            if features is None:
                display_features = ["" for i in range(len(feature_names))]
            else:
                display_features = features[i, :]

            instance = Instance(np.ones((1, len(feature_names))),
                                display_features)
            e = AdditiveExplanation(
                base_value,
                np.sum(shap_values[i, :]) + base_value, shap_values[i, :],
                None, instance, link, Model(None, out_names),
                DenseData(np.ones((1, len(feature_names))),
                          list(feature_names)))
            exps.append(e)
        return iml.visualize(exps, plot_cmap=plot_cmap)
Exemple #6
0
def visualize(shap_values,
              features=None,
              feature_names=None,
              out_names=None,
              data=None):
    """ Visualize the given SHAP values with an additive force layout. """

    # backwards compatability
    if data is not None:
        warnings.warn(
            "the 'data' parameter has been renamed to 'features' for consistency"
        )
        if features is None:
            features = data

    if type(shap_values) != np.ndarray:
        return iml.visualize(shap_values)

    # convert from a DataFrame or other types
    if str(type(features)) == "<class 'pandas.core.frame.DataFrame'>":
        if feature_names is None:
            feature_names = list(features.columns)
        features = features.as_matrix()
    elif str(type(features)) == "<class 'pandas.core.series.Series'>":
        if feature_names is None:
            feature_names = list(features.index)
        features = features.as_matrix()
    elif str(type(features)) == "list":
        if feature_names is None:
            feature_names = features
        features = None
    elif len(features.shape) == 1 and feature_names is None:
        feature_names = features
        features = None

    if len(shap_values.shape) == 1:
        shap_values = np.reshape(shap_values, (1, len(shap_values)))

    if out_names is None:
        out_names = ["output value"]

    if shap_values.shape[0] == 1:
        if feature_names is None:
            feature_names = ["" for i in range(shap_values.shape[1] - 1)]
        if features is None:
            features = ["" for i in range(len(feature_names))]
        if type(features) == np.ndarray:
            features = features.flatten()

        instance = Instance(np.zeros((1, len(feature_names))), features)
        e = AdditiveExplanation(
            shap_values[0, -1], np.sum(shap_values[0, :]), shap_values[0, :-1],
            None, instance, IdentityLink(), Model(None, out_names),
            DenseData(np.zeros((1, len(feature_names))), list(feature_names)))
        return e

    else:
        exps = []
        for i in range(shap_values.shape[0]):
            if feature_names is None:
                feature_names = ["" for i in range(shap_values.shape[1] - 1)]
            if features is None:
                display_features = ["" for i in range(len(feature_names))]
            else:
                display_features = features[i, :]

            instance = Instance(np.ones((1, len(feature_names))),
                                display_features)
            e = AdditiveExplanation(
                shap_values[i, -1], np.sum(shap_values[i, :]),
                shap_values[i, :-1], None, instance, IdentityLink(),
                Model(None, out_names),
                DenseData(np.ones((1, len(feature_names))),
                          list(feature_names)))
            exps.append(e)
        return exps
Exemple #7
0
def force_plot(shap_values, features=None, feature_names=None, out_names=None, link="identity"):
    """ Visualize the given SHAP values with an additive force layout. """

    link = iml.links.convert_to_link(link)

    if type(shap_values) != np.ndarray:
        return iml.visualize(shap_values)

    # convert from a DataFrame or other types
    if str(type(features)) == "<class 'pandas.core.frame.DataFrame'>":
        if feature_names is None:
            feature_names = list(features.columns)
        features = features.as_matrix()
    elif str(type(features)) == "<class 'pandas.core.series.Series'>":
        if feature_names is None:
            feature_names = list(features.index)
        features = features.as_matrix()
    elif str(type(features)) == "list":
        if feature_names is None:
            feature_names = features
        features = None
    elif len(features.shape) == 1 and feature_names is None:
        feature_names = features
        features = None

    if len(shap_values.shape) == 1:
        shap_values = np.reshape(shap_values, (1,len(shap_values)))

    if out_names is None:
        out_names = ["output value"]

    if shap_values.shape[0] == 1:
        if feature_names is None:
            feature_names = ["" for i in range(shap_values.shape[1]-1)]
        if features is None:
            features = ["" for i in range(len(feature_names))]
        if type(features) == np.ndarray:
            features = features.flatten()

        instance = Instance(np.zeros((1,len(feature_names))), features)
        e = AdditiveExplanation(
            shap_values[0,-1],
            np.sum(shap_values[0,:]),
            shap_values[0,:-1],
            None,
            instance,
            link,
            Model(None, out_names),
            DenseData(np.zeros((1,len(feature_names))), list(feature_names))
        )
        return e

    else:
        exps = []
        for i in range(shap_values.shape[0]):
            if feature_names is None:
                feature_names = ["" for i in range(shap_values.shape[1]-1)]
            if features is None:
                display_features = ["" for i in range(len(feature_names))]
            else:
                display_features = features[i,:]

            instance = Instance(np.ones((1,len(feature_names))), display_features)
            e = AdditiveExplanation(
                shap_values[i,-1],
                np.sum(shap_values[i,:]),
                shap_values[i,:-1],
                None,
                instance,
                link,
                Model(None, out_names),
                DenseData(np.ones((1,len(feature_names))), list(feature_names))
            )
            exps.append(e)
        return exps