

In [1]:

    
import sklearn
import sklearn.datasets
import sklearn.ensemble
import numpy as np
import lime
import lime.lime_tabular
from __future__ import print_function
np.random.seed(1)

Continuous features

Loading data, training a model

For this part, we'll use the Iris dataset, and we'll train a random forest.



In [2]:

    
iris = sklearn.datasets.load_iris()



In [3]:

    
train, test, labels_train, labels_test = sklearn.cross_validation.train_test_split(iris.data, iris.target, train_size=0.80)



In [4]:

    
rf = sklearn.ensemble.RandomForestClassifier(n_estimators=500)
rf.fit(train, labels_train)









    Out[4]:





RandomForestClassifier(bootstrap=True, class_weight=None, criterion='gini',
            max_depth=None, max_features='auto', max_leaf_nodes=None,
            min_samples_leaf=1, min_samples_split=2,
            min_weight_fraction_leaf=0.0, n_estimators=500, n_jobs=1,
            oob_score=False, random_state=None, verbose=0,
            warm_start=False)



In [5]:

    
sklearn.metrics.accuracy_score(labels_test, rf.predict(test))









    Out[5]:





0.96666666666666667

Create the explainer

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:

To scale the data, so that we can meaningfully compute distances when the attributes are not on the same scale
To sample perturbed instances - which we do by sampling from a Normal(0,1), multiplying by the std and adding back the mean.



In [6]:

    
explainer = lime.lime_tabular.LimeTabularExplainer(train, feature_names=iris.feature_names, class_names=iris.target_names, discretize_continuous=True)

Explaining an instance

Since this is a multi-class classification problem, we set the top_labels parameter, so that we only explain the top class.



In [7]:

    
i = np.random.randint(0, test.shape[0])
exp = explainer.explain_instance(test[i], rf.predict_proba, num_features=2, top_labels=1)

We now explain a single instance:



In [8]:

    
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.

Checking the local linear approximation



In [9]:

    
feature_index = lambda x: iris.feature_names.index(x)



In [10]:

    
print('Increasing petal width')
temp = test[i].copy()
print('P(setosa) before:', rf.predict_proba(temp.reshape(1,-1))[0,0])
temp[feature_index('petal width (cm)')] = 1.5
print('P(setosa) after:', rf.predict_proba(temp.reshape(1,-1))[0,0])
print ()
print('Increasing petal length')
temp = test[i].copy()
print('P(setosa) before:', rf.predict_proba(temp.reshape(1,-1))[0,0])
temp[feature_index('petal length (cm)')] = 3.5
print('P(setosa) after:', rf.predict_proba(temp.reshape(1,-1))[0,0])
print()
print('Increasing both')
temp = test[i].copy()
print('P(setosa) before:', rf.predict_proba(temp.reshape(1,-1))[0,0])
temp[feature_index('petal width (cm)')] = 1.5
temp[feature_index('petal length (cm)')] = 3.5
print('P(setosa) after:', rf.predict_proba(temp.reshape(1,-1))[0,0])









    



Increasing petal width
P(setosa) before: 1.0
P(setosa) after: 0.48

Increasing petal length
P(setosa) before: 1.0
P(setosa) after: 0.55

Increasing both
P(setosa) before: 1.0
P(setosa) after: 0.03

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:



In [11]:

    
exp.show_in_notebook(show_table=True, show_all=True)