LEM2 rule induction

Description

A python2.7 implementation of LEM2 (Learning from Examples Module, version 2): a rule induction algorithm based on rough set theory.

Usage

from lem2_classifier import LEM2Classifier

X_train, y_train, X_test, y_test = ...
lem2 = LEM2Classifier()
lem2.fit(X_train, y_train)
predictions = lem2.predict(X_test)

Example

LEM2 requires a dataset containing only categorical attributes and labeled decision values. One such dataset is the well-known playtennis dataset (included in example/). First, load the data and separate the attribute values from the decision values:

>>> import numpy as np
>>> data = np.loadtxt("example/playtennis.data", dtype='str', delimiter=',')
>>> X, y = data[:,0:-1], data[:,len(data[0])-1]

Next, fit the data using the LEM2 classifier:

>>> from lem2_classifier import LEM2Classifier
>>> lem2 = LEM2Classifier()
>>> lem2.fit(X, y)
>>> lem2.print_rules(attr_names=["outlook","temperature","humidity","wind"], class_name="play tennis")
Rule: (play tennis, yes) <- (humidity, normal), (wind, weak) [Acc. 100.0, Cov. 28.6]
Rule: (play tennis, yes) <- (outlook, overcast) [Acc. 100.0, Cov. 28.6]
Rule: (play tennis, no) <- (outlook, sunny), (humidity, high) [Acc. 100.0, Cov. 21.4]
Rule: (play tennis, yes) <- (outlook, rain), (wind, weak) [Acc. 100.0, Cov. 21.4]
Rule: (play tennis, yes) <- (outlook, sunny), (humidity, normal) [Acc. 100.0, Cov. 14.3]
Rule: (play tennis, no) <- (outlook, rain), (wind, strong) [Acc. 100.0, Cov. 14.3]

The function print_rules will print the induced rules by LEM2 in a readable form, optionally supplied with attribute names, a class name, a minimum accuracy value and a minimum coverage value. After fitting the data the induced rules can be used for classifying new instances:

>>> prediction = lem2.predict(["sunny","mild","normal","strong"])
>>> print prediction
yes

By default new instances are classified using the LERS classification method. Alternatively, it is also possible to sort the ruleset based on the accuracy and coverage of each rule and to predict an instance x by returning the decision value of the first rule which covers x (or by returning the majority class if no such rule exists):

>>> prediction = lem2.predict(["sunny","mild","normal","strong"], method="first_fit")
>>> print prediction
yes