Supervised Pattern Mining and Applications to Classification - Frequent Pattern Mining - page 417

Chapter 17
Supervised Pattern Mining and Applications
to Classification

Albrecht Zimmermann and Siegfried Nijssen

Abstract In this chapter we describe the use of patterns in the analysis of supervised
data. We survey the different settings for finding patterns as well as sets of patterns.
The pattern mining settings are categorized according to whether they include class
labels as attributes in the data or whether they partition the data based on these labels.
The pattern set mining settings are categorized along several dimensions, including
whether they perform iterative mining or post-processing, operate globally or locally,
and whether they use patterns directly or indirectly for prediction.

Keywords Rules · Classification · Subgroup discovery · Prediction · Pattern sets

1 Introduction

Although early constrained pattern mining in the form of frequent itemset mining
( FIM ) focused on an unsupervised setting, a natural extension is to apply these tech-
niques in a supervised context as well. In the supervised context, one attribute (or
sometimes a small set of attributes) is considered to be special, and we are only inter-
ested in finding relationships between this attribute and the other attributes. Whereas
this limits the patterns that will be found, it makes the analysis more targeted and
in many cases more useful. Consider for instance the context of customer defection
(churn), where one wishes to find relationships between the loyalty of customers and
other characteristics of the customers; or consider applications in cheminformatics,
where one wishes to find relationships between molecular structures and their activ-
ity: in all these cases, a targeted analysis with respect to the indicated target attribute
is likely to produce the most valuable results.