upura / featureTweakPy

Interpretable Predictions of Tree-based Ensembles via Actionable Feature Tweaking (KDD 2017)

featureTweakPy

How can prediction result be changed?

Referrence

This ipynb is inspired by the following link:
http://setten-qb.hatenablog.com/entry/2017/10/22/232016

I fixed some codes and added some explanations:

• Fix load_iris() to datasets.load_iris() at In [2]
• Fix rfc.fit(x, y) to rfc.fit(x_arr, y_arr) at In [3]
• Fix aim_label = 3 to aim_label = 2 at In [7] and [22]
• Add the usage of feature_tweaking()
• Add featureTweakPy.py to extract functions

Description

Python implementation of Interpretable Predictions of Tree-based Ensembles via Actionable Feature Tweaking (KDD 2017)
https://arxiv.org/abs/1706.06691

Requirements

• Python 3.x
  • numpy
  • pandas
  • scipy.stats

Usage

Download

git clone [e-mail]:upura/featureTweakPy.git
cd featureTweakPy

Package install

if necessary

Package import

import numpy as np
from sklearn import datasets
from sklearn.ensemble import RandomForestClassifier

from featureTweakPy import feature_tweaking

Dataset import

iris = datasets.load_iris()
x_arr = iris['data']
y_arr = iris['target']

Random Forest Prediction

rfc = RandomForestClassifier()
rfc.fit(x_arr, y_arr)

Using function()

Hyper Parameters Setting

class_labels = [0, 1, 2]
aim_label = 2
epsilon = 0.1

Cost Function Setting

def cost_func(a, b):
    return np.linalg.norm(a-b)

Sample Data for Demonstration

x = x_arr[0]
x

array([5.1, 3.5, 1.4, 0.2])

Using feature_tweaking()

x_new = feature_tweaking(rfc, x, class_labels, aim_label, epsilon, cost_func)
x_new

array([5.1       , 2.9999999 , 4.75000038, 0.90000001])