Page 129 - DMGT308_CUSTOMER_RELATIONSHIP_MANAGEMENT

Customer Relationship Management

Notes

Notes Note that one a is incorrectly covered by these rules, and more tests can be added
to exclude that a from b’s cover and include it in the a’s cover.

1R Algorithm

One of the simple approaches used to find classification rules is called 1R, as it generated a one
level decision tree. This algorithm examines the “rule that classify an object on the basis of a
single attribute”.
The basic idea is that rules are constructed to test a single attribute and branch for every value of
that attribute. For each branch, the class with the best classification is the one occurring most
often in the training data. The error rate of the rules is then determined by counting the number
of instances that do not have the majority class in the training data. Finally, the error rate for
each attribute’s rule set is evaluated, and the rule set with the minimum error rate is chosen.

A comprehensive comparative evaluation of the performance of 1R and other methods on 16
datasets (many of which were most commonly used in machine learning research) was performed.
Despite it simplicity, 1R produced surprisingly accurate rules, just a few percentage points
lower in accuracy than the decision produced by the state of the art algorithm (C4). The decision
tree produced by C4 was in most cases considerably larger than 1R’s rules, and the rules generated
by 1R were much easier to interpret. 1R therefore provides a baseline performance using a
rudimentary technique to be used before progressing to more sophisticated algorithms.

Other Algorithms
Basic covering algorithms construct rules that classify training data perfectly, that is, they tend
to over fit the training set causing insufficient generalization and difficulty for processing new
data. However, for applications in real world domains, methods for handling noisy data,
mechanisms for avoiding over fitting even on training data, and relaxation requirements of the
constraints are needed. Pruning is one of the ways of dealing with these problems, and it
approaches the problem of over fitting by learning a general concept from the training set “to
improve the prediction of unseen instance”. The concept of Reduced Error Pruning (REP) was
developed by, where some of the training examples were withheld as a test set and performance
of the rule was measured on them. Also, Incremental Reduced Error Pruning (IREP) has proven
to be efficient in handling over-fitting, and it forms the basis of RIPPER. SLIPPER (Simple
Learner with Iterative Pruning to Produce Error Reduction) uses “confidence-rated boosting to
learn an ensemble of rules.”

Applications of Rule-based Algorithms

Rule based algorithms are widely used for deriving classification rules applied in medical
sciences for diagnosing illnesses, business planning, banking government and different
disciplines of science. Particularly, covering algorithms have deep roots in machine learning.
Within data mining, covering algorithms including SWAP-1, RIPPER, and DAIRY are used in
text classification, adapted in gene expression programming for discovering classification rules.

Neural Networks

Neural Networks are analytic techniques modelled after the (hypothesized) processes of learning
in the cognitive system and the neurological functions of the brain and capable of predicting

124 LOVELY PROFESSIONAL UNIVERSITY