Jerzy Stefanowski

Algorytmy indukcji reguł decyzyjnych w odkrywaniu wiedzy
"Algorithms of rule induction for knowledge discovery"

Published as: Seria Rozprawy nr 361, Wydawnictwo Politechniki Poznańskiej, 
Poznań, 2001.

Summary
Rule induction methods for knowledge discovery are considered in this 
dissertation. Two perspectives of inducing decision rules are distinguished: 
prediction and description. Data are represented in, so called, decision tables where 
rows correspond to objects and columns correspond to attributes; each element of 
this table characterizes an object by an appropriate attribute. Attributes are divided 
into condition and decision ones. The condition attributes characterize objects 
independently of partitions induced by the decision attributes. In the problems of 
object classification, values of the decision attribute correspond to the decision 
classes and objects to examples of classification (learning examples). Decision 
rules are induced from learning examples and determine whether an object 
satisfying conditions (defined on some condition attributes) belongs to an 
appropriate decision class. If input data contain inconsistencies, rough set theory 
can be used to handle them. The dissertation discusses the current state of the 
research on rule induction and presents new algorithms. First, the method of rule 
induction specific for description perspective is proposed – an algorithm 
EXPLORE. Then, another algorithm MODLEM is introduced to handle directly 
numerical data. Two new generalizations of the rough sets theory for a case of 
incomplete decision tables, i.e. containing unknown attribute values, are proposed. 
The first generalization introduces the use of non-symmetric similarity relations in 
order to handle absent attribute values. The second generalization is based on the 
use of a valued tolerance relation and deals with missing attribute values. Further 
on, problems of discovering generalized decision rules form examples of multiple 
criteria sorting problems are considered. The rough sets approach based on the 
dominance relation – specific for multiple criteria decision problems- is here 
extended to a variable consistency approach. Two algorithms for inducing decision 
rules in both these approaches are introduced: DOMLEM and DOMApriori. The 
last methodological aspect deals with using rule induction inside multiple or hybrid 
classification systems. Within this context two proposals are presented: multiple 
classifier n2 and hybrid system integrating rule and case based learning. Finally, 
practical applications of proposed methods are discussed.

Keywords: Machine Learning, Knowledge Discovery in Databases, Classifiers, 
Rough Sets Theory, Multiple-Criteria Decision Support.