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
def kmeans(X, k, round_values=True):
""" Summarize a dataset with k mean samples weighted by the number of data points they
each represent.
Parameters
----------
X : numpy.array or pandas.DataFrame
Matrix of data samples to summarize (# samples x # features)
k : int
Number of means to use for approximation.
round_values : bool
For all i, round the ith dimension of each mean sample to match the nearest value
from X[:,i]. This ensures discrete features always get a valid value.
Returns
-------
DenseData object.
"""
group_names = [str(i) for i in range(X.shape[1])]
if str(type(X)).endswith("'pandas.core.frame.DataFrame'>"):
group_names = X.columns
X = X.values
kmeans = KMeans(n_clusters=k, random_state=0).fit(X)
if round_values:
for i in range(k):
for j in range(X.shape[1]):
ind = np.argmin(np.abs(X[:,j] - kmeans.cluster_centers_[i,j]))
kmeans.cluster_centers_[i,j] = X[ind,j]
return DenseData(kmeans.cluster_centers_, group_names, None, 1.0*np.bincount(kmeans.labels_))
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
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
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)
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)
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
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
