pyBreakDown
Python implementation of breakDown package (https://github.com/pbiecek/breakDown).
Docs: https://pybreakdown.readthedocs.io.
Requirements
Nothing fancy, just python 3.5.2+ and pip.
Installation
Install directly from github
git clone https://github.com/bondyra/pyBreakDown
cd ./pyBreakDown
python3 setup.py install # (or use pip install . instead)
Basic usage
Load dataset
from sklearn import datasets
x = datasets.load_boston()
data = x.data
feature_names = x.feature_names
y = x.target
Prepare model
import numpy as np
from sklearn import tree
model = tree.DecisionTreeRegressor()
Train model
train_data = data[1:300,:]
train_labels=y[1:300]
model = model.fit(train_data,y=train_labels)
Explain predictions on test data
#necessary imports
from pyBreakDown.explainer import Explainer
from pyBreakDown.explanation import Explanation
#make explainer object
exp = Explainer(clf=model, data=train_data, colnames=feature_names)
#make explanation object that contains all information
explanation = exp.explain(observation=data[302,:],direction="up")
Text form of explanations
#get information in text form
explanation.text()
Feature Contribution Cumulative
Intercept = 1 29.1 29.1
RM = 6.495 -1.98 27.12
TAX = 329.0 -0.2 26.92
B = 383.61 -0.12 26.79
CHAS = 0.0 -0.07 26.72
NOX = 0.433 -0.02 26.7
RAD = 7.0 0.0 26.7
INDUS = 6.09 0.01 26.71
DIS = 5.4917 -0.04 26.66
ZN = 34.0 0.01 26.67
PTRATIO = 16.1 0.04 26.71
AGE = 18.4 0.06 26.77
CRIM = 0.09266 1.33 28.11
LSTAT = 8.67 4.6 32.71
Final prediction 32.71
Baseline = 0
#customized text form
explanation.text(fwidth=40, contwidth=40, cumulwidth = 40, digits=4)
Feature Contribution Cumulative
Intercept = 1 29.1 29.1
RM = 6.495 -1.9826 27.1174
TAX = 329.0 -0.2 26.9174
B = 383.61 -0.1241 26.7933
CHAS = 0.0 -0.0686 26.7247
NOX = 0.433 -0.0241 26.7007
RAD = 7.0 0.0 26.7007
INDUS = 6.09 0.0074 26.708
DIS = 5.4917 -0.0438 26.6642
ZN = 34.0 0.0077 26.6719
PTRATIO = 16.1 0.0385 26.7104
AGE = 18.4 0.0619 26.7722
CRIM = 0.09266 1.3344 28.1067
LSTAT = 8.67 4.6037 32.7104
Final prediction 32.7104
Baseline = 0
Visual form of explanations
explanation.visualize()
#customize height, width and dpi of plot
explanation.visualize(figsize=(8,5),dpi=100)
#for different baselines than zero
explanation = exp.explain(observation=data[302,:],direction="up",useIntercept=True) # baseline==intercept
explanation.visualize(figsize=(8,5),dpi=100)
