from __future__ import print_function
import sklearn
import sklearn.datasets
import sklearn.ensemble
import pandas as pd
import numpy as np
import lime
import lime.lime_tabular
import h2o
from h2o.estimators.random_forest import H2ORandomForestEstimator
from h2o.estimators.gbm import H2OGradientBoostingEstimator
np.random.seed(1)
# Start an H2O virtual cluster that uses 6 gigs of RAM and 6 cores
h2o.init(max_mem_size = "500M", nthreads = 6)
# Clean up the cluster just in case
h2o.remove_all()
We need a wrapper class that makes an H2O distributed random forest behave like a scikit learn random forest. We instantiate the class with an h2o distributed random forest object and column names. The predict_proba method takes a numpy array as input and returns an array of predicted probabilities for each class.
class h2o_predict_proba_wrapper:
# drf is the h2o distributed random forest object, the column_names is the
# labels of the X values
def __init__(self,model,column_names):
self.model = model
self.column_names = column_names
def predict_proba(self,this_array):
# If we have just 1 row of data we need to reshape it
shape_tuple = np.shape(this_array)
if len(shape_tuple) == 1:
this_array = this_array.reshape(1, -1)
# We convert the numpy array that Lime sends to a pandas dataframe and
# convert the pandas dataframe to an h2o frame
self.pandas_df = pd.DataFrame(data = this_array,columns = self.column_names)
self.h2o_df = h2o.H2OFrame(self.pandas_df)
# Predict with the h2o drf
self.predictions = self.model.predict(self.h2o_df).as_data_frame()
# the first column is the class labels, the rest are probabilities for
# each class
self.predictions = self.predictions.iloc[:,1:].as_matrix()
return self.predictions
For this part, we'll use the Iris dataset, and we'll replace the scikit learn random forest with an h2o distributed random forest.
iris = sklearn.datasets.load_iris()
# Get the text names for the features and the class labels
feature_names = iris.feature_names
class_labels = 'species'
# Generate a train test split and convert to pandas and h2o frames
train, test, labels_train, labels_test = sklearn.model_selection.train_test_split(iris.data, iris.target, train_size=0.80)
train_h2o_df = h2o.H2OFrame(train)
train_h2o_df.set_names(iris.feature_names)
train_h2o_df['species'] = h2o.H2OFrame(iris.target_names[labels_train])
train_h2o_df['species'] = train_h2o_df['species'].asfactor()
test_h2o_df = h2o.H2OFrame(test)
test_h2o_df.set_names(iris.feature_names)
test_h2o_df['species'] = h2o.H2OFrame(iris.target_names[labels_test])
test_h2o_df['species'] = test_h2o_df['species'].asfactor()
# rf = sklearn.ensemble.RandomForestClassifier(n_estimators=500)
# rf.fit(train, labels_train)
iris_drf = H2ORandomForestEstimator(
model_id="iris_drf",
ntrees=500,
stopping_rounds=2,
score_each_iteration=True,
seed=1000000,
balance_classes=False,
histogram_type="AUTO")
iris_drf.train(x=feature_names,
y='species',
training_frame=train_h2o_df)
# sklearn.metrics.accuracy_score(labels_test, rf.predict(test))
iris_drf.model_performance(test_h2o_df)
The explainer requires numpy arrays as input and h2o requires the train and test data to be in h2o frames. In this case we could just use the train and test numpy arrays but for illustrative purposes here is how to convert an h2o frame to a pandas dataframe and a pandas dataframe to a numpy array.
train_pandas_df = train_h2o_df[feature_names].as_data_frame()
train_numpy_array = train_pandas_df.as_matrix()
test_pandas_df = test_h2o_df[feature_names].as_data_frame()
test_numpy_array = test_pandas_df.as_matrix()
As opposed to lime_text.TextExplainer, tabular explainers need a training set. The reason for this is because we compute statistics on each feature (column). If the feature is numerical, we compute the mean and std, and discretize it into quartiles. If the feature is categorical, we compute the frequency of each value. For this tutorial, we'll only look at numerical features.
We use these computed statistics for two things:
#explainer = lime.lime_tabular.LimeTabularExplainer(train, feature_names=iris.feature_names, class_names=iris.target_names, discretize_continuous=True)
explainer = lime.lime_tabular.LimeTabularExplainer(train_numpy_array,
feature_names=feature_names,
class_names=iris.target_names,
discretize_continuous=True)
We have a trained h2o distributed random forest that we would like to explain. We need to create a wrapper instance that will make it behave as a scikit learn random forest for our Lime explainer.
h2o_drf_wrapper = h2o_predict_proba_wrapper(iris_drf,feature_names)
Since this is a multi-class classification problem, we set the top_labels parameter, so that we only explain the top class.
# i = np.random.randint(0, test_numpy_array.shape[0])
i = 27
# exp = explainer.explain_instance(test[i], rf.predict_proba, num_features=2, top_labels=1)
exp = explainer.explain_instance(test_numpy_array[i], h2o_drf_wrapper.predict_proba, num_features=2, top_labels=1)
We now explain a single instance:
exp.show_in_notebook(show_table=True, show_all=False)
Now, there is a lot going on here. First, note that the row we are explained is displayed on the right side, in table format. Since we had the show_all parameter set to false, only the features used in the explanation are displayed.
The value column displays the original value for each feature.
Note that LIME has discretized the features in the explanation. This is because we let discretize_continuous=True in the constructor (this is the default). Discretized features make for more intuitive explanations.
feature_index = lambda x: iris.feature_names.index(x)
print(test[i])
print(test_numpy_array[i])
# hide the parse and predict progress bars
h2o.no_progress()
print('Increasing petal width')
temp = test_numpy_array[i].copy()
print('P(setosa) before:', h2o_drf_wrapper.predict_proba(temp)[0,0])
temp[feature_index('petal width (cm)')] = 1.5
print('P(setosa) after:', h2o_drf_wrapper.predict_proba(temp)[0,0])
print ()
print('Increasing petal length')
temp = test_numpy_array[i].copy()
print('P(setosa) before:', h2o_drf_wrapper.predict_proba(temp)[0,0])
temp[feature_index('petal length (cm)')] = 3.5
print('P(setosa) after:', h2o_drf_wrapper.predict_proba(temp)[0,0])
print()
print('Increasing both')
temp = test_numpy_array[i].copy()
print('P(setosa) before:', h2o_drf_wrapper.predict_proba(temp)[0,0])
temp[feature_index('petal width (cm)')] = 1.5
temp[feature_index('petal length (cm)')] = 3.5
print('P(setosa) after:', h2o_drf_wrapper.predict_proba(temp)[0,0])
Note that both features had the impact we thought they would. The scale at which they need to be perturbed of course depends on the scale of the feature in the training set.
We now show all features, just for completeness:
exp.show_in_notebook(show_table=True, show_all=True)
For this part, we will use the Mushroom dataset, which will be downloaded here. The task is to predict if a mushroom is edible or poisonous, based on categorical features.
# data = np.genfromtxt('http://archive.ics.uci.edu/ml/machine-learning-databases/mushroom/agaricus-lepiota.data', delimiter=',', dtype='<U20')
data = np.genfromtxt('data/mushroom_data.csv', delimiter=',', dtype='<U20')
labels = data[:,0]
le= sklearn.preprocessing.LabelEncoder()
le.fit(labels)
labels = le.transform(labels)
class_names = le.classes_
data = data[:,1:]
categorical_features = range(22)
feature_names = 'cap-shape,cap-surface,cap-color,bruises?,odor,gill-attachment,gill-spacing,gill-size,gill-color,stalk-shape,stalk-root,stalk-surface-above-ring, stalk-surface-below-ring, stalk-color-above-ring,stalk-color-below-ring,veil-type,veil-color,ring-number,ring-type,spore-print-color,population,habitat'.split(',')
We expand the characters into words, using the data available in the UCI repository
categorical_names = '''bell=b,conical=c,convex=x,flat=f,knobbed=k,sunken=s
fibrous=f,grooves=g,scaly=y,smooth=s
brown=n,buff=b,cinnamon=c,gray=g,green=r,pink=p,purple=u,red=e,white=w,yellow=y
bruises=t,no=f
almond=a,anise=l,creosote=c,fishy=y,foul=f,musty=m,none=n,pungent=p,spicy=s
attached=a,descending=d,free=f,notched=n
close=c,crowded=w,distant=d
broad=b,narrow=n
black=k,brown=n,buff=b,chocolate=h,gray=g,green=r,orange=o,pink=p,purple=u,red=e,white=w,yellow=y
enlarging=e,tapering=t
bulbous=b,club=c,cup=u,equal=e,rhizomorphs=z,rooted=r,missing=?
fibrous=f,scaly=y,silky=k,smooth=s
fibrous=f,scaly=y,silky=k,smooth=s
brown=n,buff=b,cinnamon=c,gray=g,orange=o,pink=p,red=e,white=w,yellow=y
brown=n,buff=b,cinnamon=c,gray=g,orange=o,pink=p,red=e,white=w,yellow=y
partial=p,universal=u
brown=n,orange=o,white=w,yellow=y
none=n,one=o,two=t
cobwebby=c,evanescent=e,flaring=f,large=l,none=n,pendant=p,sheathing=s,zone=z
black=k,brown=n,buff=b,chocolate=h,green=r,orange=o,purple=u,white=w,yellow=y
abundant=a,clustered=c,numerous=n,scattered=s,several=v,solitary=y
grasses=g,leaves=l,meadows=m,paths=p,urban=u,waste=w,woods=d'''.split('\n')
for j, names in enumerate(categorical_names):
values = names.split(',')
values = dict([(x.split('=')[1], x.split('=')[0]) for x in values])
data[:,j] = np.array(list(map(lambda x: values[x], data[:,j])))
Our explainer (and most classifiers) takes in numerical data, even if the features are categorical. We thus transform all of the string attributes into integers, using sklearn's LabelEncoder. We use a dict to save the correspondence between the integer values and the original strings, so that we can present this later in the explanations.
The h2o drf classifier can handle the original text, but we will convert to integers to Lime and then have h20 treat the categories as strings.
categorical_names = {}
for feature in categorical_features:
le = sklearn.preprocessing.LabelEncoder()
le.fit(data[:, feature])
data[:, feature] = le.transform(data[:, feature])
categorical_names[feature] = le.classes_
data[:,0]
categorical_names[0]
We now split the data into training and testing
data = data.astype(float)
# Generate a train test split and convert to pandas and h2o frames
train, test, labels_train, labels_test = sklearn.model_selection.train_test_split(data, labels, train_size=0.80)
class_names=['edible', 'poisonous']
train_h2o_df = h2o.H2OFrame(train)
train_h2o_df.set_names(feature_names)
train_h2o_df['class'] = h2o.H2OFrame(labels_train)
train_h2o_df['class'] = train_h2o_df['class'].asfactor()
test_h2o_df = h2o.H2OFrame(test)
test_h2o_df.set_names(feature_names)
test_h2o_df['class'] = h2o.H2OFrame(labels_test)
test_h2o_df['class'] = test_h2o_df['class'].asfactor()
Finally, we use a One-hot encoder, so that the classifier does not take our categorical features as continuous features. We will use this encoder only for the classifier, not for the explainer - and the reason is that the explainer must make sure that a categorical feature only has one value.
A great feature of h2o is that it can handle categorical data directly. There is no need to one-hot encode. We will use the integer data for the explainer and have h2o treat the integers as text.
We do have to tell h2o that the integer values in the data should be treated as categories and not numeric values. ".asfactor()" does this.
for feature in categorical_features:
train_h2o_df[feature] = train_h2o_df[feature].asfactor()
test_h2o_df[feature] = test_h2o_df[feature].asfactor()
# encoder = sklearn.preprocessing.OneHotEncoder(categorical_features=categorical_features)
# encoder.fit(data)
# encoded_train = encoder.transform(train)
# rf = sklearn.ensemble.RandomForestClassifier(n_estimators=500)
# rf.fit(encoded_train, labels_train)
# Train an h2o drf
mushroom_drf = H2ORandomForestEstimator(
model_id="mushroom_drf",
ntrees=500,
stopping_rounds=2,
score_each_iteration=True,
seed=1000000,
balance_classes=False,
histogram_type="AUTO")
mushroom_drf.train(x=feature_names,
y='class',
training_frame=train_h2o_df)
Note that our predict function first transforms the data into the one-hot representation
Create a predictor object
# predict_fn = lambda x: rf.predict_proba(encoder.transform(x))
h2o_drf_wrapper = h2o_predict_proba_wrapper(mushroom_drf,feature_names)
This classifier has perfect accuracy on the test set!
# sklearn.metrics.accuracy_score(labels_test, rf.predict(encoder.transform(test)))
mushroom_drf.model_performance(test_h2o_df)
We now create our explainer. The categorical_features parameter lets it know which features are categorical (in this case, all of them). The categorical names parameter gives a string representation of each categorical feature's numerical value, as we saw before.
np.random.seed(1)
explainer = lime.lime_tabular.LimeTabularExplainer(train ,class_names=class_names, feature_names = feature_names,
categorical_features=categorical_features,
categorical_names=categorical_names, kernel_width=3, verbose=True)
i = 137
exp = explainer.explain_instance(test[i], h2o_drf_wrapper.predict_proba, num_features=5)
exp.show_in_notebook()
Now note that the explanations are based not only on features, but on feature-value pairs. For example, we are saying that odor=foul is indicative of a poisonous mushroom. In the context of a categorical feature, odor could take many other values (see below). Since we perturb each categorical feature drawing samples according to the original training distribution, the way to interpret this is: if odor was not foul, on average, this prediction would be 0.24 less 'poisonous'. Let's check if this is the case
odor_idx = feature_names.index('odor')
explainer.categorical_names[odor_idx]
explainer.feature_frequencies[odor_idx]
foul_idx = 4
non_foul = np.delete(explainer.categorical_names[odor_idx], foul_idx)
non_foul_normalized_frequencies = explainer.feature_frequencies[odor_idx].copy()
non_foul_normalized_frequencies[foul_idx] = 0
non_foul_normalized_frequencies /= non_foul_normalized_frequencies.sum()
print('Making odor not equal foul')
temp = test[i].copy()
print('P(poisonous) before:', h2o_drf_wrapper.predict_proba(temp)[0,1])
print
average_poisonous = 0
for idx, (name, frequency) in enumerate(zip(explainer.categorical_names[odor_idx], non_foul_normalized_frequencies)):
if name == 'foul':
continue
temp[odor_idx] = idx
p_poisonous = h2o_drf_wrapper.predict_proba(temp)[0,1]
average_poisonous += p_poisonous * frequency
print('P(poisonous | odor=%s): %.2f' % (name, p_poisonous))
print ()
print ('P(poisonous | odor != foul) = %.2f' % average_poisonous)
We see that in this particular case, the linear model is pretty close: it predicted that on average odor increases the probability of poisonous by 0.26, when in fact it is by 0.23. Notice though that we only changed one feature (odor), when the linear model takes into account perturbations of all the features at once.
We now turn to a dataset that has both numerical and categorical features. Here, the task is to predict whether a person makes over 50K dollars per year. Downloads the data here.
feature_names = ["Age", "Workclass", "fnlwgt", "Education", "Education-Num", "Marital Status","Occupation", "Relationship", "Race", "Sex", "Capital Gain", "Capital Loss","Hours per week", "Country"]
# data = np.genfromtxt('http://archive.ics.uci.edu/ml/machine-learning-databases/adult/adult.data', delimiter=', ', dtype=str)
data = np.genfromtxt('data/adult.csv', delimiter=', ', dtype=str)
labels = data[:,14]
le= sklearn.preprocessing.LabelEncoder()
le.fit(labels)
labels = le.transform(labels)
class_names = le.classes_
data = data[:,:-1]
categorical_features = [1,3,5, 6,7,8,9,13]
categorical_names = {}
for feature in categorical_features:
le = sklearn.preprocessing.LabelEncoder()
le.fit(data[:, feature])
data[:, feature] = le.transform(data[:, feature])
categorical_names[feature] = le.classes_
data = data.astype(float)
# We don't need to one-hot encode from h2o
# encoder = sklearn.preprocessing.OneHotEncoder(categorical_features=categorical_features)
# Generate a train test split and convert to pandas and h2o frames
np.random.seed(1)
train, test, labels_train, labels_test = sklearn.model_selection.train_test_split(data, labels, train_size=0.80)
train_h2o_df = h2o.H2OFrame(train)
train_h2o_df.set_names(feature_names)
train_h2o_df['class'] = h2o.H2OFrame(labels_train)
train_h2o_df['class'] = train_h2o_df['class'].asfactor()
test_h2o_df = h2o.H2OFrame(test)
test_h2o_df.set_names(feature_names)
test_h2o_df['class'] = h2o.H2OFrame(labels_test)
test_h2o_df['class'] = test_h2o_df['class'].asfactor()
print(type(train_h2o_df))
# We don't need to one-hot encode from h2o
# encoder.fit(data)
# encoded_train = encoder.transform(train)
We need to tell h2o which features are categorical.
for feature in categorical_features:
train_h2o_df[feature] = train_h2o_df[feature].asfactor()
test_h2o_df[feature] = test_h2o_df[feature].asfactor()
This time, we use gradient boosted trees as the model, using the xgboost package.
We will use the H2OGradientBoostingEstimator which is h2o's version of gradient boosted trees.
adult_gbm = H2OGradientBoostingEstimator(
ntrees=300,
learn_rate=0.1,
max_depth=5,
stopping_tolerance=0.01,
stopping_rounds=2,
score_each_iteration=True,
model_id="gbm_v1",
seed=2000000
)
adult_gbm.train(feature_names, 'class', training_frame = train_h2o_df)
# sklearn.metrics.accuracy_score(labels_test, rf.predict(encoder.transform(test)))
adult_gbm.model_performance(test_h2o_df)
# predict_fn = lambda x: rf.predict_proba(encoder.transform(x)).astype(float)
h2o_drf_wrapper = h2o_predict_proba_wrapper(adult_gbm,feature_names)
explainer = lime.lime_tabular.LimeTabularExplainer(train ,feature_names = feature_names,class_names=class_names,
categorical_features=categorical_features,
categorical_names=categorical_names, kernel_width=3)
We now show a few explanations. These are just a mix of the continuous and categorical examples we showed before. For categorical features, the feature contribution is always the same as the linear model weight.
np.random.seed(1)
i = 1653
exp = explainer.explain_instance(test[i], h2o_drf_wrapper.predict_proba, num_features=5)
exp.show_in_notebook(show_all=False)
Note that capital gain has very high weight. This makes sense. Now let's see an example where the person has a capital gain below the mean:
i = 10
exp = explainer.explain_instance(test[i], h2o_drf_wrapper.predict_proba, num_features=5)
exp.show_in_notebook(show_all=False)
h2o.shutdown(prompt=False)