Graphics Reference
In-Depth Information
the minimal subset of variables needed for classification is found,
the rule is given explicitly in terms of conditions on these variables, in terms of
included and excluded intervals, and provides “a picture” that shows complex
distributions with regions where there are data and “holes” with no data, thus
providing insight to domain experts.
During - Michieetal.( ),onbehalfoftheESPRITprogramoftheEuro-
pean Union, made extensive studies of several classifiers applied to diverse datasets.
About different classifiers were applied to about datasets for comparative trials
in the StatLog project. his was designed to test classification procedures on large-
scale commercially important problems in order to determine the suitability of the
various techniques to industry. he results obtained by NC are compared with those
obtained by other well-known classifiers used in Statlog on two benchmark datasets,
and are shown in the accompanying tables.
. Satellite image dataset. his has over , items; NC's classification error was
%,k-NNwasnextwith . %,theremainingclassifiersgaveerrorsofasmuch
as %,and one was unable to classify at all. On a Vowel recognition dataset with
about , data items, NC was top with . %, next was CART-DB %, while
the rest reached down to %, with many unable to provide a classification rule,
as shown in Table . .
Rank
Classifier
Error rate %
Train
Test
NC
.
.
k - N N
.
.
L V Q
.
.
DIPOL
.
.
RBF
.
.
ALLOC
.
.
I n d C A R T
.
.
CART
.
.
Backprop
.
.
Baytree
.
.
CN
.
.
C .
.
.
NewID
.
.
Cal
.
.
Quadisc
.
.
AC
.
.
SMART
.
.
Cascade
.
.
Logdisc
.
.
Discrim
.
.
Kohonen
.
.
CASTLE
.
.
Table . . Summary of the Statlog
results and comparison with the
Nested Cavities (NC) classifier for
the satellite image data
NaiveBay
.
.
ITrule
Failed
Failed
Search WWH ::




Custom Search