%%HTML 
<iframe src="https://www.youtube.com/embed/vz_fkVkoGFM" width="560" height="315"  allowfullscreen></iframe>

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from  matplotlib.colors import LinearSegmentedColormap
gray_cmap=LinearSegmentedColormap.from_list('gy',[(.3,.3,.3),(.8,.8,.8)], N=2) 
np.random.seed(222) #Seed for random numbers generation

def set_plot_style():
  plt.axis([-2,2,-2,2])
  plt.xlabel('x1')
  plt.ylabel('x2')

df = pd.read_csv("https://arteagac.github.io/blog/lime/artificial_data.csv",header = 'infer')
df[0:5]

X = df[['x1','x2']].values
y = df['y'].values
X = (X - np.mean(X,axis=0)) / np.std(X,axis=0) #Standarization of data

set_plot_style()
plt.scatter(X[:,0],X[:,1], c=y, cmap=gray_cmap)

from sklearn.ensemble import RandomForestClassifier
classifier = RandomForestClassifier(n_estimators=100)
classifier.fit(X, y) 

#Function to create a mesh grid
def make_meshgrid(x1, x2, h=.02,x1_min=0,x1_max=0,x2_min=0,x2_max=0):
    if x1_min==0 and x1_max==0 and x2_min==0 and x2_max==0:
        x1_min, x1_max = x1.min() - 0.1, x1.max() + 0.1
        x2_min, x2_max = x2.min() - 0.1, x2.max() + 0.1
    xx1, xx2 = np.meshgrid(np.arange(x1_min, x1_max, h), np.arange(x2_min, x2_max, h))
    return np.vstack((xx1.ravel(), xx2.ravel())).T

#Create mesh grid and predict class for each element in mesh grid
XX = make_meshgrid(X[:,0],X[:,1],h=.07)
yy = classifier.predict(XX)

set_plot_style()
plt.scatter(XX[:,0],XX[:,1], c=yy, cmap=gray_cmap)

Xi = np.array([0.8,-0.7]) 
set_plot_style()
plt.scatter(XX[:,0],XX[:,1], c=yy, cmap=gray_cmap)
plt.scatter(Xi[0],Xi[1],c="blue",marker="o",s=70 )

num_perturb = 500
X_lime = np.random.normal(0,1,size=(num_perturb,X.shape[1]))

set_plot_style()
plt.scatter(X_lime[:,0],X_lime[:,1],s=2,c="black")

y_lime = classifier.predict(X_lime)
set_plot_style()
plt.scatter(X_lime[:,0],X_lime[:,1],s=5, c=y_lime, cmap=gray_cmap)

classifier.predict(np.array([0.8,-0.7]).reshape(1, -1) )


kernel_width = 0.2
distances = np.sum((Xi - X_lime)**2,axis=1) #Euclidean distance
weights = np.sqrt(np.exp(-(distances**2)/(kernel_width**2))) #Kernel function
weights.shape

set_plot_style()
plt.scatter(XX[:,0],XX[:,1], c=yy, cmap=gray_cmap) 
plt.scatter(X_lime[:,0],X_lime[:,1],s=10,c= weights,cmap="RdYlGn") 
plt.scatter(Xi[0],Xi[1],c="blue",marker="o",s=70 )

from sklearn.linear_model import LinearRegression
simpler_model = LinearRegression() 
simpler_model.fit(X_lime, y_lime, sample_weight=weights)
y_linmodel = simpler_model.predict(X_lime)
y_linmodel = y_linmodel < 0.5 #Conver to binary class

set_plot_style()
plt.scatter(XX[:,0],XX[:,1], c=yy, cmap=gray_cmap) 
plt.scatter(Xi[0],Xi[1],c="blue",marker="o",s=70 )
plt.scatter(X_lime[y_linmodel==0,0],X_lime[y_linmodel==0,1],c= weights[y_linmodel==0],cmap="RdYlGn",marker="_",s=80)
plt.scatter(X_lime[y_linmodel==1,0],X_lime[y_linmodel==1,1],c= weights[y_linmodel==1],cmap="RdYlGn",marker="+",s=80)

simpler_model.coef_