Table of Contents

• 1  Introduction
• 2  Load the libraries
• 3  Load the libraries
• 4  Useful Functions
• 5  Parameters
• 6  Load the Data
• 7  Modelling Xgboost
• 8  Using What if tool

Kernel Author:
Bhishan Poudel, Data Scientist, Ph.D Astrophysics .

Introduction

What If Tool (WIT) is developed by Google for model explanation. Here, we use the tool for xgboost model interpretation.

Load the libraries

import time
time_start_notebook = time.time()

%%capture
import os
import sys
ENV_COLAB = 'google.colab' in sys.modules

if ENV_COLAB:
    ## install modules
    !pip install watermark
    !pip install --upgrade witwidget

    # if we update existing module, we need to restart colab
    !pip install -U scikit-learn

    ## print
    print('Environment: Google Colaboratory.')
TREE_METHOD = 'gpu_hist' if ENV_COLAB else 'auto'

Load the libraries

import numpy as np
import pandas as pd
import xgboost
import sklearn
from sklearn import metrics as skmetrics

# model eval
import witwidget
from witwidget.notebook.visualization import WitConfigBuilder
from witwidget.notebook.visualization import WitWidget

# random state
SEED = 0
RNG = np.random.RandomState(SEED)

# versions
import watermark
%load_ext watermark
%watermark -a "Bhishan Poudel" -d -v -m
print()
%watermark -iv

The watermark extension is already loaded. To reload it, use:
  %reload_ext watermark
Bhishan Poudel 2020-11-21 

CPython 3.7.9
IPython 7.18.1

compiler   : Clang 10.0.0 
system     : Darwin
release    : 19.6.0
machine    : x86_64
processor  : i386
CPU cores  : 4
interpreter: 64bit

xgboost   1.2.0
sklearn   0.23.2
json      2.0.9
watermark 2.0.2
pandas    1.1.2
numpy     1.18.5

Useful Functions

def adjustedR2(rsquared,nrows,ncols):
    return rsquared- (ncols-1)/(nrows-ncols) * (1-rsquared)

def print_regr_eval(ytest,ypreds,ncols):
    rmse = np.sqrt(skmetrics.mean_squared_error(ytest,ypreds))
    r2 = skmetrics.r2_score(ytest,ypreds)
    ar2 = adjustedR2(r2,len(ytest),ncols)
    evs = skmetrics.explained_variance_score(ytest, ypreds)

    print(f"""
             RMSE : {rmse:,.2f}
Explained Variance: {evs:.6f}
         R-Squared: {r2:,.6f}
Adjusted R-squared: {ar2:,.6f}

""")

Parameters

if ENV_COLAB:
    path_git = 'https://raw.githubusercontent.com/bhishanpdl/Datasets/master/'
    project = 'Projects/King_County_Seattle_House_Price_Kaggle/'
    data_path_parent = path_git + project
else:
    data_path_parent = '../data/'


data_path_Xtest = data_path_parent + 'processed/Xtest.csv.zip'
data_path_ytest = data_path_parent + 'processed/ytest.csv'
target = 'price'
train_size = 0.8

print(data_path_Xtest)

../data/processed/Xtest.csv.zip

Load the Data

# Here, we only need test data
df_Xtest  = pd.read_csv(data_path_Xtest,compression='zip')
ser_ytest = pd.read_csv(data_path_ytest,header=None)
ytest  = np.array(ser_ytest).flatten()
features = list(df_Xtest.columns)

s = f"""
df_Xtest  = {df_Xtest.shape}
ytest     = {ytest.shape}

"""
print(s)

display(df_Xtest.head(2))
display(ser_ytest.head(2))

assert df_Xtest.shape[0] == ytest.shape[0]

df_Xtest  = (4323, 67)
ytest     = (4323,)

	age	age_after_renovation	age_after_renovation_cat	age_after_renovation_sq	age_cat	age_sq	basement_bool	bathrooms	bathrooms_sq	bedrooms	...	view_sq	waterfront	waterfront_0	waterfront_1	waterfront_sq	yr_built	yr_renovated	yr_renovated2	yr_sales	zipcode
0	-1.372335	-1.316486	-1.265291	-0.845091	-1.320662	-0.885667	-0.801818	0.506258	0.326221	-0.39033	...	-0.261712	-0.089698	0.089698	-0.089698	-0.089698	1.361464	-0.207992	1.305630	-0.693043	-1.422563
1	-0.084817	-0.005269	-0.062185	-0.285363	-0.139825	-0.348085	-0.801818	0.506258	0.326221	-0.39033	...	-0.261712	-0.089698	0.089698	-0.089698	-0.089698	0.107715	-0.207992	0.028586	1.442912	-1.441324

2 rows × 67 columns

	0
0	285000.0
1	239950.0

Modelling Xgboost

path_model_xgb = '../models/model_xgb_logtarget.dump'
model = xgboost.XGBRegressor()
model.load_model(path_model_xgb)

ypreds_log1p = model.predict(df_Xtest)
ypreds = np.expm1(ypreds_log1p)

print('ytest:', ytest[:3])
print('ypreds: ', ypreds[:3])
print_regr_eval(ytest,ypreds,df_Xtest.shape[1])

ytest: [285000. 239950. 460000.]
ypreds:  [343218.4  204292.33 508420.8 ]

             RMSE : 110,471.76
Explained Variance: 0.910365
         R-Squared: 0.909445
Adjusted R-squared: 0.908041

Using What if tool

import witwidget
from witwidget.notebook.visualization import WitConfigBuilder
from witwidget.notebook.visualization import WitWidget

def custom_predict_fn(z):
    # note: we need to use np.expm1 if we had done log1p transform if target
    testing_data = pd.DataFrame(df_Xtest, columns=df_Xtest.columns.tolist())
    return np.expm1(model.predict(testing_data))

N = 100
HEIGHT = 1000

arr_examples = np.c_[df_Xtest.to_numpy(), ytest][:N]
lst_examples = arr_examples.tolist()

config_builder = WitConfigBuilder(lst_examples, features + [target])

config_builder.set_target_feature(target)
config_builder.set_custom_predict_fn(custom_predict_fn)
config_builder.set_model_type('regression')

WitWidget(config_builder, height=HEIGHT)

!ls images/wh*

images/what_if_tool_snap.png