Database Reference
In-Depth Information
Normal
pattern
R
max
Novel
Parameter 1
Fig. 2.9.
Principle of OCC by NOVCLASS.
su
cient covering of the normal domain. In this iterative training case, a
new hypersphere with center
x
l
and radius
R
max
is added to the
initially empty classifier iff a presented vector
t
J
+1
=
x
l
from the training set is
classified as novel by the already stored
J
reference vectors
t
j
according to
the basic steps given earlier. The denseness of the NOVCLASS model poten-
tially can be controlled by scaling the
R
max
parameter by a scale factor
η
to
η × R
max
in the training process. A large-scale factor implies few stored
vectors and potential coarse window modeling, whereas a small-scale factor
η<
1 means fine window modeling at the cost of storing and processing a
potentially large number of vectors. A functional nonparametric classifier is
thus achieved with examples of just one class. Additionally, if at least a few
examples for anomalies are available, these can be used to fine-tune the radii
of the stored normal class hyperspheres by applying RCE-like adaptation for
the conflicting hyperspheres. In this case, radii will no longer be uniform.
Currently, a prototype NOVCLASS version has been implemented and
validated with modified Iris data, where all examples of class 3 were a
l-
iated to class 2. Class 1 was chosen as the normal class. Resubstitution of
the training set was perfect and in generalization just one vector slightly
separated from the main cluster was misclassified.
Summarizing, the NOVCLASS algorithm allows both data reduction and
arbitrary coverage of the parameter space. Thus, the concept of the process
window is generalized to arbitrary shapes, including no convex boundaries.
The current rather ad hoc uniform
R
max
computation approach could be
improved by more sophisticated methods, e.g., locally adaptive radii com-
putation, in future work. The present NOVCLASS implementation will be
Search WWH ::
Custom Search