Table of Contents
Abstract
Introduction
Technologies
Data Warehouse
Examples
Characteristics
Platforms
Summary
|
|
3.0 Data Mining Examples
Data mining technology has evolved in a bottom-up, application-driven
manner.
This section describes two important data mining technologies: mining for
association rules and profile generation.
3.1 Mining for Association Rules
Association rule technology has many applications, including supermarket
inventory planning, shelf planning, and attached mailing in direct
marketing.
For example, association rules can be derived from mining "market basket"
data
for buying patterns exhibited by customers. Here, each "market basket"
consists
of the set of items bought by a customer on a single visit to a store. An
example association rule in this context is:
80% of the people who buy diapers and baby powder buy baby oil.
The number "80%" is referred to as the confidence factor, a measure of the
predictive power of the rule. Here "diapers" and "baby powder" constitute
the
set of items on the left hand side (LHS) of the rule, and "baby oil" is
the
item on the right hand side (RHS) of the rule. Let us look at some
applications
of such discovered rules.
Rules that have "Diet Coke" in the RHS
Use: What should the store do to boost the sale of Diet Coke?
Rules that have "Bagels" in the LHS
Use: What products may be impacted if the store discontinues selling
Bagels?
Rules that have "Sausage" in the LHS and "Mustard" in the RHS
Use: What items should be sold with sausage to promote the sales of
mustard?
There can be any number of items on the LHS and RHS of association rules.
The
general model for association rules can be described as follows: Given a
database of transactions, where each transaction consists of a number of
items,
such as retail transactions, a pattern is a particular set of items, such
as
{Milk Bread Eggs}. We can derive candidate rules from patterns. For
example,
given the pattern {Milk Bread Eggs}, a possible rule is:
Milk & Bread ==> Eggs
The prevalence of a pattern (and derived rules) is the number of
transactions
in which the pattern appears. The prevalence of the pattern {Milk Bread
Eggs}
is denoted by:
{Milk Bread Eggs}
We are interested only in patterns and rules that have some given minimum
prevalence for two reasons.
1. A rule has business value only if a significant fraction of
transactions
support it. For example, it may be the case that everybody who buys
"Caviar"
buys "Vodka". Thus, the rule
Caviar ==> Vodka
has 100% confidence. But if only a handful of people buy "Caviar", the
rule may
be of limited value to the retailer. 2. A rule may not be statistically
significant if a very small number of transactions support the rule.
The rule may be observed due to chance, and it would not be prudent to
make
decisions based on such a rule.
The confidence factor described earlier is mathematically expressed as the
ratio of the prevalence of the rule pattern to the prevalence of the LHS
of the
rule. For the rule
Milk & Bread ==> Eggs
the confidence factor is given by
|Milk Bread Eggs|
------------------
|Milk Bread|
The techniques needed to generate these association rules automatically
require
significant I/O bandwidth and computational power, and the data sets are
large
and need to be scanned fast. So for acceptable response time, the
operations
need to be parallelized, and the hardware and software must support these
kinds
of processing.
3.2 Profile Generation
Profile generation has many applications, including target marketing,
attached
mailings, credit approval and treatment-appropriateness determination.
Consider target marketing. Company X wishes to send out a promotional
mailing.
Company X can buy a general mailing list and would like to maximize the
return
on its mailing expenditure. Profile generation can help as follows. If the
company has done a prior mailing, the responses to that mailing are
available.
The people on the previous mailing, their attributes, and their responses
are
input to the profile generator. The profile generator comes up with a
characterization, or profile, of the people who responded to the previous
mailing. This profile is then taken as a predictor of response to the
current mailing. The mailing list is filtered to include only people who match the
profile and the promotional mailing is sent to them.
As another example, consider credit or loan approval. Here, the company
wishes
to characterize the people who should be studied carefully before a loan
or
credit is granted to them. The input to the profile generation program is
the
prior lending or credit experience of the company. As with generating
association rules, profile generation techniques can benefit from
computational
power and parallelism. In addition, large main memories can help to speed
the
process up significantly.
|
|
