def LIME_graphSHAP(X, y, feature_names, ylabels, clf, xs_toexplain, labels_toexplain, ax=None, subplots=False, plotlime=True):
## Plot explanations on feature space
if ax is None:
if subplots:
if len(xs_toexplain)>=5:
nrows = int(len(xs_toexplain)/5)
fig, axs = plt.subplots(nrows=nrows, ncols=int(len(xs_toexplain)/nrows), figsize=(15,3*nrows))
axs = axs.flatten()
else:
fig, axs = plt.subplots(nrows=1, ncols=len(xs_toexplain), figsize=(15,3))
else:
fig, ax = plt.subplots()
# Plot LIME result - loop if several
for i in range(len(xs_toexplain)):
if subplots:
plt.sca(axs[i])
ax = axs[i]
if i==0 or subplots:
# Plot contour of black-box predictions
plot_classification_contour(X, clf, ax)
# Plot training set
plot_training_set(X, y, ax)
ylim_bak = ax.get_ylim()
xlim_bak = ax.get_xlim()
#color_palette = sns.color_palette("bright", n_colors=len(xs_toexplain))
color_palette = ['lime' for _ in range(len(xs_toexplain))]
## LIME - Generate explanations
explainer = lime_assessment.lime_tabular.LimeTabularExplainer(X, feature_names=feature_names, class_names=ylabels, discretize_continuous=False, kernel_width=None)
exp = explainer.explain_instance(xs_toexplain[i], clf.predict_proba, num_features=2, top_labels=len(ylabels), labels=range(len(ylabels)))
shap_explainer = KernelExplainer(clf.predict_proba, X, nsamples=10000)
e = shap_explainer.explain(np.reshape(xs_toexplain[i], (1, X.shape[1])))
# Plot LIME regression
if plotlime == True:
plot_lime_regression(X, exp, xs_toexplain[i], labels_toexplain[i], ax, color_palette[i], exp.points_to_plot)
x_ridge = [-10, 10]
row = 0
y_shap = [(0.5 - e.effects[0, row] * x - e.base_value[row])/e.effects[1, row] for x in x_ridge]
# Plot LIME linear regression
plt.sca(ax)
plt.plot(x_ridge, y_shap, color='red', linestyle=':', linewidth=4, label="other shap regression")
plt.scatter(xs_toexplain[i][0], xs_toexplain[i][1], color='lime', marker='8', linewidth=4)
plt.ylim(ylim_bak)
plt.xlim(xlim_bak)
def test_front_page_model_agnostic():
from shap import KernelExplainer, DenseData, visualize, initjs
from sklearn import datasets,neighbors
from numpy import random, arange
# print the JS visualization code to the notebook
initjs()
# train a k-nearest neighbors classifier on a random subset
iris = datasets.load_iris()
random.seed(2)
inds = arange(len(iris.target))
random.shuffle(inds)
knn = neighbors.KNeighborsClassifier()
knn.fit(iris.data, iris.target == 0)
# use Shap to explain a single prediction
background = DenseData(iris.data[inds[:100],:], iris.feature_names) # name the features
explainer = KernelExplainer(knn.predict, background, nsamples=100)#knn.predict
x = iris.data[inds[102:103],:]
visualize(explainer.explain(x))
