Extracting Local Model Coef_

Hi, this is just a followup issue with regards to the local model weights.

I have the following code.

from sklearn.linear_model import LogisticRegression

clf = LogisticRegression()
clf.fit(X_train, y_train)  # Just the Taiwan credit dataset

explainer = lime.lime_tabular.LimeTabularExplainer(X_train, 
                                                   feature_names=feature_names, 
                                                   class_names=['N', "Y"],
                                                   verbose=False, 
                                                   discretize_continuous=False,
                                                   mode='classification')

i = 13
exp = explainer.explain_instance(X_train[i], clf.predict_proba)

total = 0
l = exp.as_map()[1]
l.sort()

for j in range(len(X_train[0])):
    total += X_train[i][j] * (l[j][1])
total + exp.intercept[1]

The issue is that my manual calculation of exp.local_pred is never exactly correct. Am I doing something wrong here? It should just be a simple matter of multiplying the input features by the weights plus intercept no?

The training data is just 3 continuous features.

Thanks a lot in advance!

[EoinKenny]

@marcotcr you were exactly right, it works perfect now.

Thanks a lot for taking the time to explain.

[marcotcr]

That is not normal. I think there are two bugs in your code:

1. X_train[i, j] will be a categorical value (an integer) if feature[j] is categorical, whereas the explanation weight should be multiplied by 1, indicating the presence of the categorical value.
2. coefs will give you at most num_feature coefs (this is a parameter to explain_instance, the default is 10). If you have more than 10 features in your dataset, the multiplication won’t work.

I think if you do

ids = [x[0] for x in coefs]
scaled_instance[categorical_features] = 1
manual_calculation = sum(coefs * scaled_instance[ids]) + exp.intercept[1]

you will get the same as exp.local_pred