Database Reference
In-Depth Information
classes—different classes have different products of probabilities and the class with
the highest probability is predicted.
This is somewhat similar to the Krimp algorithm: in this technique,
coding tables
are created for each class separately, and an unseen instance's label is predicted based
on the coding table that compresses it best.
These approaches are arguably still limited by what the pattern themselves can
do, although the upshot is that their models are somewhat more understandable. The
alternative is to mine patterns as features for use in sophisticated machine learning
techniques that can add modeling and generalization capabilities that are missing
from symbolic patterns themselves. This is the second big group of techniques: the
technique proposed by [
22
] belongs to it, as does DDPMINE, the method introduced
by [
12
], PICKER
∗
, FCORK, and REMINE.
5
Summary
In this chapter, we have given a high level overview of supervised pattern mining
and its application to prediction, specifically classification. We have abstracted from
the pattern languages used and structured the chapter along the three main steps
involved in building a classifier from class-labeled data: supervised pattern mining,
supervised pattern set mining, and classifier construction.
Regarding the first step, we have laid out that many techniques view different
classes as separate subsets of the data and evaluate patterns' co-occurrence with one
of these subsets. In our opinion, this view clarifies that different quality measures
will lead to similar semantical information of patterns, and that different mining
approaches can be taken to find patterns that score highly with any of these measures.
Regarding the second step, we have pointed out the similarities to approaches that
have been pioneered in machine learning in the context of rule learning, decision
tree induction, and instance-based learning. We have interpreted the former two
approaches in terms of partitions to show the similarities of existing techniques,
and also identified two types of approaches that always manipulate the entire data.
Although some pattern set mining techniques, in particular iterative ones, make
certain demands on the pattern mining step, most of them can still be combined
relatively freely with different pattern mining techniques.
Finally, when it comes to classifier building, we have made the distinction between
direct and indirect classification, with the former paralleling rule-based classification
in machine learning, and the latter comprising quite a few approaches that mine
patterns as
features
for use in propositional learners. As a comparison of references
shows, different classifiers also do not track closely with particular pattern or pattern
set mining approaches.
In general, in surveying the field we find that many solutions to the three phases
have been developed, most of which can be mixed-and-matched rather freely. The
field is larger than the algorithms we have mentioned here yet many techniques are
arguably variations of the approaches that we have contrasted.
