LIME local prediction computed from weights of features

lime_local_prediction_manual_calculation.py

import numpy as np
import pandas as pd
import xgboost as xgb

from sklearn.model_selection import train_test_split
from sklearn.datasets import make_regression

import lime.lime_tabular

### make data
X, y = make_regression(n_samples=100, n_features=5,  n_informative=3, random_state=0, noise=4.0, bias=10.0)
feature_names = ["x" + str(i+1) for i in range(0,5)]
data = pd.DataFrame(X, columns=feature_names)
data["target"] = y
X_train, X_test, y_train, y_test = train_test_split(data[feature_names], ## predictors only
                                                    data.target,
                                                    test_size=0.30, 
                                                    random_state=0)

### create and fit model
estimator = xgb.XGBRegressor()
estimator.fit(X_train, y_train)

## receive data as numpy array which has no column names, so we fix it
def xgb_predict(data_asarray):
    data_asframe =  pd.DataFrame(data_asarray, columns=feature_names)
    return estimator.predict(data_asframe)

instance = X_test.iloc[2]
exp = lime_tabular_explainer.explain_instance(instance, xgb_predict, num_features=5)

## instance should be scaled as coefs are create from a scaled data set, see https://github.com/marcotcr/lime/issues/189
scaled_instance = (instance - lime_tabular_explainer.scaler.mean_) / lime_tabular_explainer.scaler.scale_

local_exp = exp.as_map()[1]
coefs = [x[1] for x in local_exp]
ids = [x[0] for x in local_exp]

print("Manual calculation of local pred:", sum(coefs * scaled_instance[ids]) + exp.intercept[1])