# -*- coding: utf-8 -*-

"""

@author: dgarreau

In this script, we show how Tabular LIME ignores unused coordinates for a

non-linear f. We take the example of a kernel ridge regressor trained on a

subset of the coordinates of the Wine dataset. This is Figure 15 in the paper.

"""

import numpy as np

import matplotlib.pyplot as plt

import lime

import lime.lime_tabular

from sklearn.datasets import load_wine

from sklearn.model_selection import train_test_split

from sklearn import preprocessing

from sklearn.kernel_ridge import KernelRidge

from sklearn.gaussian_process.kernels import RBF

from utils.aux_functions import get_training_data_stats

from utils.aux_functions import format_coefs

from utils.plot_functions import plot_whisker_boxes

if __name__ == "__main__":

# number of experiments

n_exp = 100

# number of perturbed samples

n_samples = 1000

# for reproducibility

np.random.seed(1)

# non linear model, train some kernel ridge regressor on Wine

wine = load_wine()

X_orig = wine['data']

# removing the last n_removed coordinate

n_removed = 6

X = X_orig[:,:-n_removed]

Y = wine['target']

# scale evrything

X_scaled = preprocessing.scale(X)

X_orig_scaled = preprocessing.scale(X_orig)

# train test split

X_train, X_test, Y_train, Y_test = train_test_split(X_scaled, Y, test_size = 0.2, random_state=5)

# let us train the model

kernel = RBF(length_scale=10)

krr_model = KernelRidge(kernel=kernel,alpha=1)

krr_model.fit(X_train,Y_train)

# and restrict to the first coordinates

def my_model(array):

return krr_model.predict(array[:,:-n_removed])

# dimension of the ambient space

dim = X_orig.shape[1]

# the example to explain

xi = X_orig_scaled[0,:]

# bandwidth parameter

nu = 5

# number of bins along each dimension

p = 4

# getting the stats

my_stats = get_training_data_stats(X_orig_scaled,p)

# creating the explainer

explainer = lime.lime_tabular.LimeTabularExplainer(X_orig_scaled,

mode='regression',

feature_selection='none',

training_data_stats=my_stats,

kernel_width=nu)

# main loop

beta_emp_store = np.zeros((n_exp,dim+1))

for i_exp in range(n_exp):

if np.mod(i_exp + 1,10) == 0:

s_exp = "Experiment {} / {} is running...".format(i_exp + 1,n_exp)

print(s_exp)

# getting the explanation

exp = explainer.explain_instance(xi,

my_model,

num_samples=n_samples)

# getting the coefficients of the local model

beta_emp_store[i_exp,:] = format_coefs(exp)

###########################################################################

# getting nicer feature names

#my_features = wine['feature_names']

my_features = ['alcohol','Malic acid','ash','alcalinity','Mg',

'phenols','flavanoids','non-flavanoid phenol',

'proanthocyanins','color intensity','hue',

'part of diluted wines','proline']

# plotting the result

fig, ax = plt.subplots(figsize=(15,10))

plot_whisker_boxes(beta_emp_store,

ax,

rotate=True,

feature_names=my_features)

s_name = "results/ignore_non_linear_default_weights"

fig.savefig(s_name + '.pdf',format='pdf',bbox_inches = 'tight',pad_inches = 0)