def decisions_regression(
df_preds,
shap_values,
expected_value,
X_vald,
y_vald,
model_file_path,
learner_name,
):
fig = plt.gcf()
shap.decision_plot(
expected_value,
shap_values[df_preds.lp[:10], :],
X_vald.loc[df_preds.index[:10]],
show=False,
)
fig.tight_layout(pad=2.0)
fig.savefig(
os.path.join(model_file_path,
f"{learner_name}_shap_worst_decisions.png"))
plt.close("all")
fig = plt.gcf()
shap.decision_plot(
expected_value,
shap_values[df_preds.lp[-10:], :],
X_vald.loc[df_preds.index[-10:]],
show=False,
)
fig.tight_layout(pad=2.0)
fig.savefig(
os.path.join(model_file_path,
f"{learner_name}_shap_best_decisions.png"))
plt.close("all")
def explain_row_shap(scaled_row, explainer, nsamples=100, verbose=0):
shap_values = explainer.shap_values(scaled_row.reshape(1, shape_size),
nsamples=nsamples,
l1_reg="num_features(32)")
if (verbose == 1):
shap.decision_plot(explainer.expected_value[0],
shap_values[0][0, :],
scaled_row,
feature_names=list(
df.drop('loan_repaid', axis=1).columns),
link="logit")
map_values = {}
for class_value in range(len(shap_values)):
s = shap_values[class_value][0]
sorted_indices = sorted(range(len(s)),
key=lambda k: s[k],
reverse=True)
# print(sorted_indices)
ordered_list = [(a, shap_values[class_value][0][a])
for a in sorted_indices
if shap_values[class_value][0][a] > 0]
map_values[class_value] = ordered_list
# print(map_values)
return map_values
def get_decision_plot(patente, step_id_week):
f = plt.figure()
ranker = joblib.load(
'C:/Users/raskolnnikov/Desktop/projects/samtech/samtech_entrega/modelos/ranker_v_1.0.joblib'
)
explainer = shap.TreeExplainer(ranker)
ranker_features = ranker.feature_names
expected_value = explainer.expected_value
entry = RankingEntry.query.filter_by(
patente=patente, step_id_week=step_id_week).one().instance_json
instance = pd.read_json(entry, orient='records').T
instance.columns = ranker_features
shap_value = ranker.predict(
xgb.DMatrix(instance[ranker_features]),
pred_contribs=True,
)[0, :-1]
shap.decision_plot(expected_value,
shap_value,
instance.iloc[0, :],
link='logit',
highlight=0,
show=False)
buf = BytesIO()
f.savefig(buf, format="png", dpi=150, bbox_inches='tight')
buf.seek(0)
f.clear()
plt.close(f)
img_base64 = base64.b64encode(buf.read())
response = {
"image": img_base64.decode(),
}
return response, 200
def shap_decision_plot(expected_value,
shap_values,
samples,
save=True,
crop_feature_names=20):
if crop_feature_names:
feature_names = []
for col in samples.columns:
if len(col) > crop_feature_names:
feature_names.append(col[:20])
else:
feature_names.append(col)
shap.decision_plot(expected_value,
shap_values,
features=samples,
feature_names=feature_names,
show=False)
else:
shap.decision_plot(expected_value,
shap_values,
features=samples,
show=False)
f = plt.gcf()
f.show()
if save:
f.savefig("shap_decision_plot.png")
def combine_summary_decision_curve(shap_value, expected_value, features,
feature_names, n_features,
examples_subset_index, misclassified, link,
save_path):
"""
Generate a combined SHAP Summary and Decision Plot
Parameters
----------
shap_value: np.ndarray
SHAP value for a particular output class
expected_value: float
Average of the classifier/model output over training dataset
features: np.ndarray
Features in testing dataset
feature_names: np.array
List of feature names
n_features: int
Maximum count of features to use for the plots
examples_subset_index: list of indices
Samples to select for the decision plot
misclassified: list of Boolean values
Denotes which selected samples are misclassified (set to true)
link: string
Link type to be used for the decision plot
save_path: string
File path where the plot will be saved
"""
figsize = (12, 6)
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=figsize)
# Generate Summary plot
plt.sca(ax1)
shap.summary_plot(shap_value,
features=features,
feature_names=feature_names,
sort=True,
show=False,
max_display=n_features)
# Generate Decision plot
plt.gcf().set_size_inches(figsize)
plt.sca(ax2)
feature_order = np.argsort(np.sum(np.abs(shap_value), axis=0))
shap.decision_plot(expected_value,
shap_value[examples_subset_index],
features[examples_subset_index, :],
feature_names=list(feature_names),
feature_display_range=slice(None, -(n_features + 1),
-1),
ignore_warnings=True,
highlight=misclassified,
show=False,
plot_color='viridis',
feature_order=feature_order,
link=link)
plt.plot([0.5, 0.5], [0, n_features], ':k', alpha=0.3)
ax2.set_yticklabels([])
plt.savefig(save_path)
plt.close()
def decision_plot(self, class_id=0, row_idx=-1, **kwargs):
"Visualize model decision using cumulative `SHAP` values."
shap_vals, exp_val = _get_values(self, class_id)
n_rows = shap_vals.shape[0]
if row_idx == -1:
print(f'Displaying rows 0-9 of {n_rows} (use `row_idx` to specify another row)')
return shap.decision_plot(exp_val, shap_vals[:10], self.test_data.iloc[:10], **kwargs)
print(f'Displaying row {row_idx} of {n_rows} (use `row_idx` to specify another row)')
return shap.decision_plot(exp_val, shap_vals[row_idx], self.test_data.iloc[row_idx], **kwargs)
def plot_shap_force(drug_id, expected_value, shap_values_test, data_for_shap,
drug_names, X_train, force_plot_file_type, dpi,
eval_label):
curr_shap_value = shap_values_test.loc[
drug_id, :] # explainer.shap_values(curr_drug_features)
curr_features = data_for_shap.loc[drug_id, :]
drug_name = drug_names.loc[drug_id, 'Drug name']
title = 'Probability higher risk (%s)' % (drug_name)
curr_feature_vals = np.array([
'Yes' if bool(x) else
"No" if X_train.columns[i] != 'Number of Category' else x
for i, x in enumerate(curr_features.values)
])
p = shap.force_plot(
base_value=expected_value,
shap_values=curr_shap_value.values,
#feature_names=[x.replace("Cluster: ",'').replace(';','\n') for x in X_train.columns], #x.split(': ')[1] if ': ' in x else
feature_names=[
x.split(': ')[1] if ': ' in x else x for x in X_train.columns
],
features=
curr_feature_vals, #['Yes' if x else 'No' for x in curr_features.values],
out_names=[title],
figsize=(20, 4) #
,
show=False,
matplotlib=True,
text_rotation=int(45 / 2))
p.savefig(os.path.join('output', 'SHAP' + "_" + eval_label,
drug_id + '.' + force_plot_file_type),
dpi=dpi,
bbox_inches='tight')
# p.show()
plt.close('all')
shap.decision_plot(
base_value=expected_value,
shap_values=curr_shap_value.values,
feature_names=[
x.split(': ')[1] if ': ' in x else x for x in X_train.columns
],
features=curr_feature_vals,
feature_display_range=slice(-1, -11, -1),
title=title,
show=False,
#link='logit',
#highlight=0
)
p = plt.gcf()
p.savefig(os.path.join('output', 'SHAP' + "_" + eval_label,
drug_id + '_decision_plot.' + force_plot_file_type),
dpi=dpi,
bbox_inches='tight')
plt.close('all')
def decision_plot(self, X, y):
"""Visualization of the additive feature attribution."""
# Automates single-target slicing
y = super()._check_target_index(y=y)
for index in range(_n_targets(y)):
self.fit(X=X, y=y, index=index)
explainer, shap_values = self.explainer(X=X)
shap.decision_plot(base_value=explainer.expected_value,
shap_values=shap_values,
feature_names=list(X.columns),
show=self.show)
def visualise_explanation(explanation, per_class=True):
"""
Visualises an explanation of classification performance.
Parameters
----------
explanation
Output of `classification.explain_classifier()`.
per_class : bool, optional
Whether to also plot explanations at the level of the individual
classes, or just the summary.
"""
# Summarise across all classes
fig, ax = plt.subplots()
shap.summary_plot(explanation.shap_values,
explanation.data,
class_names=explanation.clf_categories,
max_display=15)
ax.set_ylabel("Feature")
ax.set_title("Classifier feature weightings (all classes)")
fig.tight_layout()
# And then break down by class
if per_class:
for i in range(len(explanation.shap_values)):
# Summary plot: break down by feature
fig, ax = plt.subplots()
shap.summary_plot(explanation.shap_values[i],
explanation.data,
max_display=10)
ax.set_ylabel("Feature")
ax.set_title("Classifier feature weightings (class: {})".format(
explanation.clf_categories[i]))
fig.tight_layout()
# Decision plot: break down by observation
fig, ax = plt.subplots()
shap.decision_plot(explanation.explainer.expected_value[i],
explanation.shap_values[i],
link='logit',
features=explanation.data)
ax.set_ylabel("Feature")
ax.set_title("Classifier decision weightings (class: {})".format(
explanation.clf_categories[i]))
fig.tight_layout()
return
def shap_explain(booster,
datasource,
dataset,
summary_params,
result_table="",
is_pai=False,
oss_dest=None,
oss_ak=None,
oss_sk=None,
oss_endpoint=None,
oss_bucket_name=None):
tree_explainer = shap.TreeExplainer(booster)
shap_values = tree_explainer.shap_values(dataset)
if result_table:
if is_pai:
conn = PaiIOConnection.from_table(result_table)
else:
conn = db.connect_with_data_source(datasource)
# TODO(typhoonzero): the shap_values is may be a
# list of shape [3, num_samples, num_features],
# use the first dimension here, should find out
# when to use the other two. When shap_values is
# not a list it can be directly used.
if isinstance(shap_values, list):
to_write = shap_values[0]
else:
to_write = shap_values
columns = list(dataset.columns)
with db.buffered_db_writer(conn, result_table, columns) as w:
for row in to_write:
w.write(list(row))
conn.close()
if summary_params.get("plot_type") == "decision":
shap_interaction_values = tree_explainer.shap_interaction_values(
dataset)
expected_value = tree_explainer.expected_value
if isinstance(shap_interaction_values, list):
shap_interaction_values = shap_interaction_values[0]
if isinstance(expected_value, list):
expected_value = expected_value[0]
plot_func = lambda: shap.decision_plot( # noqa: E731
expected_value,
shap_interaction_values,
dataset,
show=False,
feature_display_range=slice(None, -40, -1),
alpha=1)
else:
plot_func = lambda: shap.summary_plot( # noqa: E731
shap_values, dataset, show=False, **summary_params)
explainer.plot_and_save(plot_func,
oss_dest=oss_dest,
oss_ak=oss_ak,
oss_sk=oss_sk,
oss_endpoint=oss_endpoint,
oss_bucket_name=oss_bucket_name,
filename='summary')
def decision_plot(self,
fold,
supersample_frac=.1,
feature_display_range=None,
cmap=None):
"""
For a small sample of test data points, starting from the models base prediction value,
incrementely adding features to each prediction and plotting the path of the model score
as a continuous line. Good for visualising how features across the model combine to create
a diverse range of model scores.
"""
expected_value = self.fold_expected_values[fold]
shap_values = self.shapley_values_array[:, :, fold]
X = self.SHAP_X_sample
super_sample = X.sample(frac=supersample_frac, replace=False)
super_sample_idx = super_sample.reset_index().index.values
super_sampled_shapley_values = shap_values[super_sample_idx, :]
return shap.decision_plot(expected_value,
super_sampled_shapley_values,
super_sample,
feature_display_range=feature_display_range,
plot_color=cmap)
def local_plot(name,
explainer,
shap_values,
feature_names,
chosen_sample,
estimand_name,
X_test,
plot_type='force_plot'):
if plot_type == 'force_plot':
h = shap.force_plot(base_value=explainer.expected_value,
shap_values=shap_values[chosen_sample],
features=X_test[chosen_sample],
feature_names=feature_names,
link="identity",
out_names=estimand_name,
matplotlib=False,
show=False)
save_plot(h, name)
elif plot_type == 'decision_plot':
h = shap.decision_plot(base_value=explainer.expected_value,
shap_values=shap_values[chosen_sample],
features=X_test[chosen_sample],
feature_names=feature_names,
link="identity",
matplotlib=False,
show=False)
save_plot(h, name)
return h
def decision_plot(self, X, y):
"""Visualization of the additive feature attribution."""
y = self._slice_target_index(y=y)
for index in range(_n_targets(y)):
if sklearn.utils.multiclass.type_of_target(
y) == 'continuous-multioutput':
self.fit(X, y.iloc[:, index].values.ravel(order='K'))
else:
self.fit(X, y)
explainer, shap_values = self.explainer(X=X)
shap.decision_plot(base_value=explainer.expected_value,
shap_values=shap_values,
feature_names=list(X.columns),
show=self.show)
def shap_explain(datasource,
select,
feature_field_meta,
feature_column_names,
label_meta,
summary_params,
result_table="",
is_pai=False,
pai_explain_table="",
oss_dest=None,
oss_ak=None,
oss_sk=None,
oss_endpoint=None,
oss_bucket_name=None,
transform_fn=None,
feature_column_code=""):
x = xgb_shap_dataset(datasource,
select,
feature_column_names,
label_meta,
feature_field_meta,
is_pai,
pai_explain_table,
transform_fn=transform_fn,
feature_column_code=feature_column_code)
shap_values, shap_interaction_values, expected_value = xgb_shap_values(x)
if result_table != "":
if is_pai:
from runtime.dbapi.paiio import PaiIOConnection
conn = PaiIOConnection.from_table(result_table)
else:
conn = db.connect_with_data_source(datasource)
# TODO(typhoonzero): the shap_values is may be a
# list of shape [3, num_samples, num_features],
# use the first dimension here, should find out
# when to use the other two. When shap_values is
# not a list it can be directly used.
if isinstance(shap_values, list):
to_write = shap_values[0]
else:
to_write = shap_values
write_shap_values(to_write, conn, result_table, feature_column_names)
if summary_params.get("plot_type") == "decision":
explainer.plot_and_save(
lambda: shap.decision_plot(expected_value,
shap_interaction_values,
x,
show=False,
feature_display_range=slice(
None, -40, -1),
alpha=1), oss_dest, oss_ak, oss_sk,
oss_endpoint, oss_bucket_name)
else:
explainer.plot_and_save(
lambda: shap.summary_plot(
shap_values, x, show=False, **summary_params), oss_dest,
oss_ak, oss_sk, oss_endpoint, oss_bucket_name)
def shap_explain(booster, datasource, dataset, summary_params, result_table):
tree_explainer = shap.TreeExplainer(booster)
shap_values = tree_explainer.shap_values(dataset)
if result_table:
conn = db.connect_with_data_source(datasource)
# TODO(typhoonzero): the shap_values is may be a
# list of shape [3, num_samples, num_features],
# use the first dimension here, should find out
# when to use the other two. When shap_values is
# not a list it can be directly used.
if isinstance(shap_values, list):
to_write = shap_values[0]
else:
to_write = shap_values
columns = list(dataset.columns)
dtypes = [DataType.to_db_field_type(conn.driver, DataType.FLOAT32)
] * len(columns)
_create_table(conn, result_table, columns, dtypes)
with db.buffered_db_writer(conn, result_table, columns) as w:
for row in to_write:
w.write(list(row))
conn.close()
if summary_params.get("plot_type") == "decision":
shap_interaction_values = tree_explainer.shap_interaction_values(
dataset)
expected_value = tree_explainer.expected_value
if isinstance(shap_interaction_values, list):
shap_interaction_values = shap_interaction_values[0]
if isinstance(expected_value, list):
expected_value = expected_value[0]
plot_func = lambda: shap.decision_plot( # noqa: E731
expected_value,
shap_interaction_values,
dataset,
show=False,
feature_display_range=slice(None, -40, -1),
alpha=1)
else:
plot_func = lambda: shap.summary_plot( # noqa: E731
shap_values, dataset, show=False, **summary_params)
filename = 'summary.png'
with temp_file.TemporaryDirectory(as_cwd=True):
explainer.plot_and_save(plot_func, filename=filename)
with open(filename, 'rb') as f:
img = f.read()
img = base64.b64encode(img)
if six.PY3:
img = img.decode('utf-8')
img = "<div align='center'><img src='data:image/png;base64,%s' /></div>" \
% img
print(img)
def decision_plot(
self, num_samples=0.25, sample_no=None, output_file="", **decisionplot_kwargs
):
"""
Plots a SHAP decision plot.
Parameters
----------
num_samples : int, float, or 'all', optional
Number of samples to display, if less than 1 it will treat it as a percentage, 'all' will include all samples
, by default 0.25
sample_no : int, optional
Sample number to isolate and analyze, if provided it overrides num_samples, by default None
Returns
-------
DecisionPlotResult
If return_objects=True (the default). Returns None otherwise.
"""
return_objects = decisionplot_kwargs.pop("return_objects", True)
highlight = decisionplot_kwargs.pop("highlight", None)
if sample_no is not None:
if sample_no < 1 or not isinstance(sample_no, int):
raise ValueError("Sample number must be greater than 1.")
samples = slice(sample_no - 1, sample_no)
else:
if num_samples == "all":
samples = slice(0, len(self.x_test_array))
elif num_samples <= 0:
raise ValueError(
"Number of samples must be greater than 0. If it is less than 1, it will be treated as a percentage."
)
elif num_samples > 0 and num_samples < 1:
samples = slice(0, int(num_samples * len(self.x_test_array)))
else:
samples = slice(0, num_samples)
if highlight is not None:
highlight = highlight[samples]
s = shap.decision_plot(
self.expected_value,
self.shap_values[samples],
self.x_train.columns,
return_objects=return_objects,
highlight=highlight,
show=False,
**decisionplot_kwargs,
)
if output_file: # pragma: no cover
pl.savefig(os.path.join(IMAGE_DIR, self.model_name, output_file))
return s
def get_shap(model, exp_data, i):
shap.initjs()
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(exp_data)
force_plot = shap.force_plot(explainer.expected_value[i],
shap_values[i][0, :], exp_data.iloc[0, :])
decision_plot = shap.decision_plot(explainer.expected_value[i],
shap_values[i][0, :],
exp_data.iloc[0, :])
return force_plot
def explain(self, options, instance=None):
if instance is None:
raise ValueError("Instance was not provided")
initjs()
instance = instance.to_numpy()
data = self._kmeans(options['kmeans_count']) \
if options['background_data'] == 'kmeans' else options['data']
nsamples = 'auto' if options['auto_nsamples'] else options['nsamples']
explainer = KernelExplainer(model=self.predict_function,
data=data,
link=options['link'])
shap_values = explainer.shap_values(X=instance,
nsamples=nsamples,
l1_reg=options['l1_reg'])
if self.is_classification:
shap_values = shap_values[options['class_to_explain']]
base_value = explainer.expected_value[[
options['class_to_explain']
]]
else:
base_value = explainer.expected_value
if options['plot_type'] == 'force' or options['plot_type'] == 'both':
display(
force_plot(base_value=base_value,
shap_values=shap_values,
features=instance,
feature_names=self.feature_names,
show=True,
link=options['link']))
if options['plot_type'] == 'decision' or options['plot_type'] == 'both':
decision_plot(base_value=base_value,
shap_values=shap_values,
features=instance,
feature_names=list(self.feature_names),
show=True,
color_bar=True,
link=options['link'])
def decisions_binary(
df_preds,
shap_values,
expected_value,
X_vald,
y_vald,
model_file_path,
learner_name,
):
# classes are from 0 ...
for t in np.unique(y_vald):
fig = plt.gcf()
shap.decision_plot(
expected_value,
shap_values[df_preds[df_preds.target == t].lp[:10], :],
X_vald.loc[df_preds[df_preds.target == t].index[:10]],
show=False,
)
fig.tight_layout(pad=2.0)
fig.savefig(
os.path.join(
model_file_path,
f"{learner_name}_shap_class_{t}_worst_decisions.png",
)
)
plt.close("all")
fig = plt.gcf()
shap.decision_plot(
expected_value,
shap_values[df_preds[df_preds.target == t].lp[-10:], :],
X_vald.loc[df_preds[df_preds.target == t].index[-10:]],
show=False,
)
fig.tight_layout(pad=2.0)
fig.savefig(
os.path.join(
model_file_path, f"{learner_name}_shap_class_{t}_best_decisions.png"
)
)
plt.close("all")
def explain(datasource,
select,
feature_field_meta,
feature_column_names,
label_spec,
summary_params,
result_table="",
is_pai=False,
pai_explain_table="",
hdfs_namenode_addr="",
hive_location="",
hdfs_user="",
hdfs_pass="",
oss_dest=None,
oss_ak=None,
oss_sk=None,
oss_endpoint=None,
oss_bucket_name=None):
x = xgb_shap_dataset(datasource, select, feature_column_names, label_spec,
feature_field_meta, is_pai, pai_explain_table)
shap_values, shap_interaction_values, expected_value = xgb_shap_values(x)
if result_table != "":
if is_pai:
# TODO(typhoonzero): the shape of shap_values is (3, num_samples, num_features)
# use the first dimension here, should find out how to use the other two.
write_shap_values(shap_values[0], "pai_maxcompute", None,
result_table, feature_column_names,
hdfs_namenode_addr, hive_location, hdfs_user,
hdfs_pass)
else:
conn = connect_with_data_source(datasource)
write_shap_values(shap_values[0], conn.driver, conn, result_table,
feature_column_names, hdfs_namenode_addr,
hive_location, hdfs_user, hdfs_pass)
return
if summary_params.get("plot_type") == "decision":
explainer.plot_and_save(
lambda: shap.decision_plot(expected_value,
shap_interaction_values,
x,
show=False,
feature_display_range=slice(
None, -40, -1),
alpha=1), is_pai, oss_dest, oss_ak,
oss_sk, oss_endpoint, oss_bucket_name)
else:
explainer.plot_and_save(
lambda: shap.summary_plot(
shap_values, x, show=False, **summary_params), is_pai,
oss_dest, oss_ak, oss_sk, oss_endpoint, oss_bucket_name)
def shap_plots(model, train_features, test_features, test_labels):
print("Computing shapley values..")
# compute SHAP values
if isinstance(
model,
(MLP, MLPRegressor, MLPClassifier, ElasticNet, LogisticRegression)):
train_sample = shap.sample(train_features, 10)
explainer = shap.Explainer(model.predict, train_sample)
elif isinstance(model, (RandomForestRegressor, RandomForestClassifier)):
explainer = shap.TreeExplainer(model, train_features)
else:
explainer = shap.Explainer(model, train_features)
shap_values = explainer(test_features)
shap.plots.bar(shap_values, max_display=10)
# shap.plots.bar(shap_values[0]) # Local
# beeswarm plot
shap.plots.beeswarm(shap_values)
# Decision plot
expected_value = explainer.expected_value
select = range(20)
features_sample = test_features.iloc[select]
shap.decision_plot(expected_value, explainer.shap_values(features_sample),
features_sample)
# Heatmap
shap.plots.heatmap(shap_values, max_display=10)
# Scatter
shap.plots.scatter(shap_values[:, "hs_child_age_None"],
color=shap_values,
alpha=0.8)
# Feature clustering (redondant feature detection)
clustering = shap.utils.hclust(
test_features, test_labels
) # by default this trains (X.shape[1] choose 2) 2-feature XGBoost models
shap.plots.bar(shap_values, clustering=clustering, clustering_cutoff=0.5)
def decision_plot(self, class_id=0, **kwargs):
"""
Visualize model decisions using cumulative SHAP values. Each colored line in the plot represents the model
prediction for a single observation. Note that plotting too many samples at once can make the plot unintelligible.
`class_id` is used to indicate the class of interest for classification models, it can ba an int or a string
For an up-to-date list of the parameters, see: https://github.com/slundberg/shap/blob/master/shap/plots/decision.py
For more informations, see: https://github.com/slundberg/shap/blob/master/notebooks/plots/decision_plot.ipynb
"""
# NOTE: there is a shap.multioutput_decision_plot but it uses a single row
shap_values, expected_value = _get_values(self, class_id)
return shap.decision_plot(expected_value, shap_values, self.test_data, **kwargs)
def explain_one_sample(self, X):
'''
Draws decision plot for one sample
@X => one sample dataframe
'''
if X.shape[0] > 1:
raise Exception(
'You need to pass only one sample of data for this function.\nIt means sample size (1, n_features)'
)
explainer = shap.TreeExplainer(self.model[0]['clf'])
shap_values = explainer.shap_values(self.process_data(X))[1]
try:
shap.decision_plot(
explainer.expected_value[1],
shap_values,
ignore_warnings=False,
feature_names=self.model[0]['teach_cols'].tolist())
except IndexError:
shap.decision_plot(
explainer.expected_value,
shap_values,
ignore_warnings=False,
feature_names=self.model[0]['teach_cols'].tolist())
def _plot_decision_(self,
expected_value: float,
shap_values: List[np.ndarray] or np.ndarray,
title: str = None,
gene_names: bool = True,
auto_size_plot: bool = True,
minimum: int = 0.0,
maximum: int = 0.0,
feature_display_range=None,
save: Path = None):
#shap.summary_plot(shap_values, self.partitions.X, show=False)
feature_names = None if gene_names is False else self.feature_names
min_max = (self.partitions.data.y.min(), self.partitions.data.y.max())
print(f"min_max dataset values: {min_max}")
xlim = (min(min_max[0], minimum), max(min_max[1], maximum))
shap.decision_plot(expected_value,
shap_values,
xlim=xlim,
feature_names=feature_names.tolist(),
title=title,
auto_size_plot=auto_size_plot,
feature_display_range=feature_display_range,
show=False)
return self.make_figure(save)
def explain(datasource, select, feature_field_meta, label_spec,
summary_params):
feature_column_names = [k["name"] for k in feature_field_meta]
feature_specs = {k['name']: k for k in feature_field_meta}
x = xgb_shap_dataset(datasource, select, feature_column_names, label_spec,
feature_specs)
shap_values, shap_interaction_values, expected_value = xgb_shap_values(x)
if summary_params.get("plot_type") == "decision":
explainer.plot_and_save(lambda: shap.decision_plot(
expected_value,
shap_interaction_values,
x,
show=False,
feature_display_range=slice(None, -40, -1),
alpha=1))
else:
explainer.plot_and_save(lambda: shap.summary_plot(
shap_values, x, show=False, **summary_params))
""")
explainer, shap_values = load_shap_explainer(
root + 'shap_treeExplainer.bz2', X_test_transformed)
st_shap(
shap.force_plot(explainer.expected_value,
shap_values[0, :],
X_test.iloc[0, :],
link='logit'))
# -----------
shap.decision_plot(base_value=explainer.expected_value,
shap_values=shap_values[0],
features=X_test.iloc[0, :],
link='logit',
feature_display_range=slice(
None, -X_test.shape[1] - 1, -1),
return_objects=True,
show=False,
y_demarc_color='#00172b')
fig = plt.gcf()
ax = plt.gca()
fig.patch.set_facecolor('#00172b')
ax.set_facecolor('#00172b')
ax.set_xlabel('Probability', fontsize=16, color='white')
ax.tick_params(axis='both', colors='white')
ax.grid(axis='both', color='white', linestyle='-', linewidth=0.25)
for ln in ax.lines:
ln.set_linewidth(3)
for text in ax.texts:
text.set_color('white')
Next, we use the SHAP values to build up 2D scatter graphs for every feature. They shows the effect of a feature for the prediction for every instance.
fig, axs = plt.subplots(7,3,figsize=(16,22),squeeze=True)
ind = 0
for ax in axs.flat:
feat = bst.feature_names[ind]
ax.scatter(x_df[feat],shap_values_XGB_test[:,ind],s=1,color='gray')
# ax.set_ylim([-0.2,0.2])
ax.set_title(feat)
ind+=1
plt.subplots_adjust(hspace=0.8)
plt.savefig('shap_sc.png')
**Decision_plot()** is interesting as it shows how the prediction is formed from the contributions of different features.
shap.decision_plot(explainerXGB.expected_value,shap_values_XGB_test[0:100],features)
**Force_plot** is similar to decision_plot. We plot only the first 100 instances because it would be very slow to draw a force_plot with all the instances.
shap.force_plot(explainerXGB.expected_value,shap_values_XGB_test[0:100],features,figsize=(20,10))
**Waterfall_plot** is great when you want to analyse one instance.
shap.waterfall_plot(explainerXGB.expected_value,shap_values_XGB_test[2000],x_df.iloc[2000],features)
### Other interpretation methods
For the following methods, we need to use the Xgboost's Scikit-learn wrapper **XGBRegressor()** to make our Xgboost model to be compatible with the Scikit-learn ecosystem.
m_depth = 5
for mode in [df_features_ec_season, df_features_ec_season_permuted]:
shap_values = explainer.shap_values(mode,
approximate=True,
check_additivity=True)
# dependence plots
for name in mode.columns:
shap.dependence_plot(name, shap_values[1], mode)
# Summary plots
shap.summary_plot(shap_values, mode, plot_type="bar")
shap.summary_plot(shap_values[1], mode, plot_type="bar")
shap.summary_plot(shap_values[1], mode) # Failure
# Decision plots explaining decisions to classify
shap.decision_plot(explainer.expected_value[1], shap_values[1], mode)
shap.decision_plot(explainer.expected_value[1], shap_values[1][1],
mode.iloc[1]) #2012 year
# Calculate force plot for a given value 2012
shap.initjs()
shap_values_2012 = explainer.shap_values(mode.iloc[[4]])
shap_display = shap.force_plot(explainer.expected_value[1],
shap_values_2012[1],
mode.iloc[[4]],
matplotlib=True)
#%% Predictions for 2C degree
y_pred_2C = brf_model.predict(df_features_ec_2C_season)
score_prc_2C = sum(y_pred_2C) / len(y_pred_2C)
print("The ratio of failure seasons by total seasons for 2C is:", score_prc_2C)
shap.dependence_plot("user_level", shap_values, X_test)
shap.dependence_plot("user_level", shap_values, X_test, interaction_index=None)
shap.dependence_plot("education", shap_values, X_test)
shap.dependence_plot("education", shap_values, X_test, interaction_index=None)
# load JS visualization code to notebook
shap.initjs()
plt.clf()
shap.force_plot(explainer.expected_value,
shap_values[1594, :],
X_test_disp.iloc[1594, :],
matplotlib=True,
show=False,
figsize=(20, 4))
plt.savefig("prediction1.png", bbox_inches='tight')
plt.clf()
shap.force_plot(explainer.expected_value,
shap_values[1594, :],
X_test_disp.iloc[1594, :],
matplotlib=True,
show=False,
figsize=(20, 4))
plt.savefig("prediction1.eps", bbox_inches='tight', format='eps')
shap.decision_plot(explainer.expected_value, shap_values[1594, :],
X_test_disp.iloc[1594, :])
def app():
st.markdown("""<style>.big-font {font-size:100px !important;}</style>""", unsafe_allow_html=True)
st.markdown(
"""<style>
.boxBorder {
border: 2px solid #990066;
padding: 10px;
outline: #990066 solid 5px;
outline-offset: 5px;
font-size:25px;
}</style>
""", unsafe_allow_html=True)
st.markdown('<div class="boxBorder"><font color="RED">Disclaimer: This predictive tool is only for research purposes</font></div>', unsafe_allow_html=True)
st.write("## Model Perturbation Analysis")
@st.cache(hash_funcs={"MyUnhashableClass": lambda _: None})
def load_model2():
with open('saved_models/trainXGB_class_map.pkl', 'rb') as f:
class_names = list(pickle.load(f))
return class_names
class_names = load_model2()
# @st.cache(hash_funcs={"MyUnhashableClass": lambda _: None})
def load_model():
M_dict = {}
for classname in class_names:
M_dict[classname] = joblib.load('saved_models/trainXGB_gpu_{}.model'.format(classname))
return M_dict
M_dict = load_model()
@st.cache(hash_funcs={"MyUnhashableClass": lambda _: None})
def load_model1():
with open('saved_models/trainXGB_gpu_{}.data'.format(class_names[0]), 'rb') as f:
train = pickle.load(f)
with open('saved_models/trainXGB_categorical_map.pkl', 'rb') as f:
col_dict_map = pickle.load(f)
return train, col_dict_map
train, col_dict_map = load_model1()
X = train[1]['X_valid'].copy()
ids = list(train[3]['ID_test'])
X.index = ids
labels_pred = list(train[3]['y_pred_test'])
labels_actual = list(train[3]['y_test'])
# select_patient = st.selectbox("Select the patient", list(X.index), index=0)
categorical_columns = []
numerical_columns = []
X_new = X.fillna('X')
for col in X_new.columns:
# if len(X_new[col].value_counts()) <= 10:
if col_dict_map.get(col, None) is not None:
categorical_columns.append(col)
else:
numerical_columns.append(col)
st.write('### Please enter the following {} factors to perform prediction or select a random patient'.format(len(categorical_columns + numerical_columns)))
# st.write("***Categorical Columns:***", categorical_columns)
# st.write("***Numerical Columns:***", numerical_columns)
from collections import defaultdict
if st.button("Random Patient"):
import random
select_patient = random.choice(list(X.index))
else:
select_patient = list(X.index)[0]
select_patient_index = ids.index(select_patient)
new_feature_input = defaultdict(list)
for key, val in col_dict_map.items():
rval = {j:i for i,j in val.items()}
X_new[key] = X_new[key].map(lambda x: rval.get(x, x))
st.write('--'*10)
st.write('##### Note: X denoted NA values')
col1, col2, col3, col4 = st.beta_columns(4)
for i in range(0, len(categorical_columns), 4):
with col1:
if (i+0) >= len(categorical_columns):
continue
c1 = categorical_columns[i+0]
idx = list(X_new[c1].unique()).index(X_new.loc[select_patient, c1])
f1 = st.selectbox("{}".format(feature_mapping[c1]), list(X_new[c1].unique()), index=idx)
new_feature_input[c1].append(col_dict_map[c1].get(f1, np.nan))
with col2:
if (i+1) >= len(categorical_columns):
continue
c2 = categorical_columns[i+1]
idx = list(X_new[c2].unique()).index(X_new.loc[select_patient, c2])
f2 = st.selectbox("{}".format(feature_mapping[c2]), list(X_new[c2].unique()), index=idx)
new_feature_input[c2].append(col_dict_map[c2].get(f2, np.nan))
with col3:
if (i+2) >= len(categorical_columns):
continue
c3 = categorical_columns[i+2]
idx = list(X_new[c3].unique()).index(X_new.loc[select_patient, c3])
f3 = st.selectbox("{}".format(feature_mapping[c3]), list(X_new[c3].unique()), index=idx)
new_feature_input[c3].append(col_dict_map[c3].get(f3, np.nan))
with col4:
if (i+3) >= len(categorical_columns):
continue
c4 = categorical_columns[i+3]
idx = list(X_new[c4].unique()).index(X_new.loc[select_patient, c4])
f4 = st.selectbox("{}".format(feature_mapping[c4]), list(X_new[c4].unique()), index=idx)
new_feature_input[c4].append(col_dict_map[c4].get(f4, np.nan))
for col in numerical_columns:
X_new[col] = X_new[col].map(lambda x: float(x) if not x=='X' else np.nan)
for i in range(0, len(numerical_columns), 4):
with col1:
if (i+0) >= len(numerical_columns):
continue
c1 = numerical_columns[i+0]
idx = X_new.loc[select_patient, c1]
f1 = st.number_input("{}".format(feature_mapping[c1]), min_value=X_new[c1].min(), max_value=X_new[c1].max(), value=idx)
new_feature_input[c1].append(f1)
with col2:
if (i+1) >= len(numerical_columns):
continue
c2 = numerical_columns[i+1]
idx = X_new.loc[select_patient, c2]
f2 = st.number_input("{}".format(feature_mapping[c2]), min_value=X_new[c2].min(), max_value=X_new[c2].max(), value=idx)
new_feature_input[c2].append(f2)
with col3:
if (i+2) >= len(numerical_columns):
continue
c3 = numerical_columns[i+2]
idx = X_new.loc[select_patient, c3]
f3 = st.number_input("{}".format(feature_mapping[c3]), min_value=X_new[c3].min(), max_value=X_new[c3].max(), value=idx)
new_feature_input[c3].append(f3)
with col4:
if (i+3) >= len(numerical_columns):
continue
c4 = numerical_columns[i+3]
idx = X_new.loc[select_patient, c4]
f4 = st.number_input("{}".format(feature_mapping[c4]), min_value=X_new[c4].min(), max_value=X_new[c4].max(), value=idx)
new_feature_input[c4].append(f4)
st.write('--'*10)
st.write("### Do you want to see the effect of changing a factor on this patient?")
color_discrete_map = {}
color_discrete_map_list = ["red", "green", "blue", "goldenred", "magenta", "yellow", "pink", "grey"]
for e, classname in enumerate(class_names):
color_discrete_map[classname] = color_discrete_map_list[e]
show_whatif = st.checkbox("Enable what-if analysis")
col01, col02 = st.beta_columns(2)
with col01:
st.write('### Prediction on actual feature values')
feature_print = X_new.loc[select_patient, :].fillna('X')
feature_print.index = feature_print.index.map(lambda x: feature_mapping[x])
feature_print = feature_print.reset_index()
feature_print.columns = ["Feature Name", "Feature Value"]
st.table(feature_print.set_index("Feature Name").astype(str))
predicted_prob = defaultdict(list)
predicted_class = -1
max_val = -1
for key, val in M_dict.items():
predicted_prob['predicted_probability'].append(val.predict(xgb.DMatrix(X.loc[select_patient, :].values.reshape(1, -1), feature_names=X.columns))[0])
predicted_prob['classname'].append(key)
if predicted_prob['predicted_probability'][-1] > max_val:
predicted_class = key
max_val = predicted_prob['predicted_probability'][-1]
K = pd.DataFrame(predicted_prob)
K['predicted_probability'] = K['predicted_probability'] / K['predicted_probability'].sum()
K['color'] = ['zed' if i==predicted_class else 'red' for i in list(predicted_prob['classname']) ]
# fig = px.bar(K, x='predicted_probability', y='classname', color='color', width=500, height=400, orientation='h')
# # fig = px.bar(K, y='predicted_probability', x=sorted(list(predicted_prob['classname'])), width=500, height=400)
# fig.update_layout(
# legend=None,
# yaxis_title="Class Labels",
# xaxis_title="Predicted Probability",
# font=dict(
# family="Courier New, monospace",
# size=12,
# color="RebeccaPurple"
# ),
# margin=dict(l=10, r=10, t=10, b=10),
# )
# st.plotly_chart(fig)
import altair as alt
K = K.rename(columns={"classname": "Class Labels", "predicted_probability": "Predicted Probability"})
f = alt.Chart(K).mark_bar().encode(
y=alt.Y('Class Labels:N',sort=alt.EncodingSortField(field="Predicted Probability", order='descending')),
x=alt.X('Predicted Probability:Q'),
color=alt.Color('color', legend=None),
).properties(width=500, height=300)
st.write(f)
# st.write('#### Trajectory for Predicted Class')
st.write('#### Model Output Trajectory for {} Class using SHAP values'.format(predicted_class))
@st.cache(hash_funcs={"MyUnhashableClass": lambda _: None})
def load_model5():
with open('saved_models/trainXGB_gpu_{}.data'.format(predicted_class), 'rb') as f:
new_train = pickle.load(f)
return new_train
new_train = load_model5()
exval = new_train[2]['explainer_train']
explainer_train = shap.TreeExplainer(M_dict[predicted_class])
t1 = pd.DataFrame(X.loc[select_patient, :]).T
t2 = pd.DataFrame(X_new.loc[select_patient, :].fillna('X')).T
shap_values_train = explainer_train.shap_values(t1)
shap.force_plot(exval, shap_values_train, t1, show=False, matplotlib=True)
st.pyplot()
fig, ax = plt.subplots()
r = shap.decision_plot(exval, shap_values_train, t2, link='logit', return_objects=True, new_base_value=0, highlight=0)
st.pyplot(fig)
if show_whatif:
with col02:
dfl = pd.DataFrame(new_feature_input)
ndfl = dfl.copy()
for key, val in col_dict_map.items():
rval = {j:i for i,j in val.items()}
ndfl[key] = ndfl[key].map(lambda x: rval.get(x, x))
st.write('### Prediction with what-if analysis')
feature_print_what = ndfl.iloc[0].fillna('X')
feature_print_what.index = feature_print_what.index.map(lambda x: feature_mapping[x])
feature_print_what = feature_print_what.reset_index()
feature_print_what.columns = ["Feature Name", "Feature Value"]
selected = []
for i in range(len(feature_print_what)):
if feature_print.iloc[i]["Feature Value"] == feature_print_what.iloc[i]["Feature Value"]:
pass
else:
selected.append(feature_print.iloc[i]["Feature Name"])
# st.table(feature_print)
st.table(feature_print_what.astype(str).set_index("Feature Name").style.apply(lambda x: ['background: yellow' if (x.name in selected) else 'background: lightgreen' for i in x], axis=1))
dfl = dfl[X.columns].replace('X', np.nan)
predicted_prob = defaultdict(list)
predicted_class = -1
max_val = -1
for key, val in M_dict.items():
predicted_prob['predicted_probability'].append(val.predict(xgb.DMatrix(dfl.iloc[0, :].values.reshape(1, -1), feature_names=dfl.columns))[0])
predicted_prob['classname'].append(key)
if predicted_prob['predicted_probability'][-1] > max_val:
predicted_class = key
max_val = predicted_prob['predicted_probability'][-1]
K = pd.DataFrame(predicted_prob)
K['predicted_probability'] = K['predicted_probability'] / K['predicted_probability'].sum()
K['color'] = ['zed' if i==predicted_class else 'red' for i in list(predicted_prob['classname']) ]
import altair as alt
K = K.rename(columns={"classname": "Class Labels", "predicted_probability": "Predicted Probability"})
f = alt.Chart(K).mark_bar().encode(
y=alt.Y('Class Labels:N',sort=alt.EncodingSortField(field="Predicted Probability", order='descending')),
x=alt.X('Predicted Probability:Q'),
color=alt.Color('color', legend=None),
).properties( width=500, height=300)
st.write(f)
# fig = px.bar(K, x='predicted_probability', y='classname', color='color', width=500, height=400, orientation='h')
# # fig = px.bar(K, y='predicted_probability', x=sorted(list(predicted_prob['classname'])), width=500, height=400)
# fig.update_layout(
# legend=None,
# yaxis_title="Class Labels",
# xaxis_title="Predicted Probability",
# font=dict(
# family="Courier New, monospace",
# size=12,
# color="RebeccaPurple"
# ),
# margin=dict(l=10, r=10, t=10, b=10),
# )
# st.plotly_chart(fig)
st.write('#### Model Output Trajectory for {} Class using SHAP values'.format(predicted_class))
@st.cache(hash_funcs={"MyUnhashableClass": lambda _: None})
def load_model6():
with open('saved_models/trainXGB_gpu_{}.data'.format(predicted_class), 'rb') as f:
new_train = pickle.load(f)
return new_train
new_train = load_model6()
exval = new_train[2]['explainer_train']
explainer_train = shap.TreeExplainer(M_dict[predicted_class])
t1 = dfl.copy()
shap_values_train = explainer_train.shap_values(t1)
shap.force_plot(exval, shap_values_train, t1, show=False, matplotlib=True)
st.pyplot()
fig, ax = plt.subplots()
_ = shap.decision_plot(exval, shap_values_train, ndfl.fillna('X'), link='logit', feature_order=r.feature_idx, return_objects=True, new_base_value=0, highlight=0)
st.pyplot(fig)
