Companion Rule List

Python code for Companion rule list.

Requirements

To run this program, we need common libs such as numpy ro pandas.
Especially, bitarray and fim are necessary to run the program.

lib list:

• python 3.6
• bitarray
• fim
• numpy
• pandas
• itertools
• operator
• scipy
• maptlotlib
• math
• argparse

Usage:

Prepare data:

Data should be prepared in binary form.
e.g.

age	gender	y
20	0	0
30	1	1

Set Parameters:

• --file: binarized data, last column should be the label, default as adult.csv
• --blx_file: black box classification label, default as rf_adult.csv
• --alpha: hyper parameter to control rule length, default as 0.001
• --step: training steps, default as 20000
• --card: rule cardinality, default as 2
• --supp: rule mining min support, keep rules cover > support, default as 0.05
• --n: number of positive and negtive rules, defaule as 200

Rusult form

The output form of the program is shown as follows:
[if condition_1 is 1/0 and condition_2 is 1/0, then y= 1/0. transparency = 20%, accuracy = 90%]
[if condition_3 is 1/0 and condition_4 is 1/0, then y= 1/0. transparency = 40%, accuracy = 88%]
...

• condition_1 and condition_2 represents the condition for making decision.
• Transparency represents the proportion of data classified by the rule list.
• Accuracy represents the prediction performance of hybrid models with corresponding number of rules.