Logic-Based Pattern Discovery

[attachment=17369]
INTRODUCTION
THE use of association rule mining technique is to
describe the associations among items in a database.
These associations represent the domain knowledge encapsulated
in databases. Identifying domain knowledge is
important because these knowledge rules usually are
known only by the domain experts over years of experience.
Thus, association rule mining is useful to identify domain
knowledge hidden in large volume of data efficiently.
The discovery of association rules is typically based on
the support and confidence framework where a minimum
support (min sup) must be supplied to start the discovery
process [1]. A priori is a representational algorithm based on
this framework and many other algorithms are a priori-like.
Without this threshold specified, typically, no association
rules can be discovered because the procedure to discover
the rules will quickly exhaust the available resources.



ISSUES USING A MINIMUM SUPPORT THRESHOLD
Issues with discovering association rules reverberate
around loss of rules and quality of rules discovered.
Specifically, if rules are lost, it is misleading to report an
incomplete set of rules and at the same time create a sense
that all available rules have been found. This situation
misleads a decision maker into thinking that only these
rules are available which, in turn, will lead a decision maker
to reason with incomplete information. For example, it is
erroneous to assume that a subset of an incomplete set of
rules has the strongest rules. Reasoning with incomplete
information while not knowing it may lead to inappropriate
conclusion or decisions.


Loss of Association Rules Involving Infrequently Observed Items

Some infrequent association rules are actionable. Typically,
a data set contains items that appear frequently while other
items rarely occur. For example, in a retail fruit business,
fruits are frequently observed but occasionally bread is also
observed. Some items are rare in nature or infrequently
found in a data set. These items are called rare items [7], [8],
[9]. If a single minimum support threshold is used and is set
high, those association rules involving rare items will not be
discovered. Use of a single and lower minimum support
threshold, on the other hand, would result in too many
uninteresting association rules. This is called the rare item
problem defined by Mannila [10] according to Liu et al. [7].