Weighting of Features by Sequential Selection - Feature Selection for Data and Pattern Recognition - page 81

Table 5.2 Sequential backward elimination of attributes basing on the performance of ANN
classifiers
(a) (b) (c) (d) (e)
0 25 but and not in with on at of as this that by for to if what from . , ; : ! ? ( - 83.33 not
1 24 but and in with on at of as this that by for to if what from . , ; : ! ? ( - 88.33 (
2 23 but and in with on at of as this that by for to if what from . , ; : ! ? - 90.00 in
3 22 but and with on at of as this that by for to if what from . , ; : ! ? - 91.67 and
4 21 butwithonatofasthisthatbyfortoifwhatfrom.,;:!?- 93.33 ,
5 20 but with on at of as this that by for to if what from . ; : ! ? - 94.17 -
6 19 but with on at of as this that by for to if what from . ; : ! ? 95.00 with
7 18 but on at of as this that by for to if what from . ; : ! ? 96.67 on
8 17 but at of as this that by for to if what from . ; : ! ? 96.67 what
9 16 but at of as this that by for to if from . ; : ! ? 96.67 to
10 15 but at of as this that by for if from . ; : ! ? 96.67 of
11 14 but at as this that by for if from . ; : ! ? 96.67 .
12 13 but at as this that by for if from ; : ! ? 96.67 :
13 12 but at as this that by for if from ; ! ? 96.67 ?
14 11 but at as this that by for if from ; ! 96.67 ;
15 10 but at as this that by for if from ! 96.67 this
16 9 but at as that by for if from ! 95.00 as
17 8 but at that by for if from ! 95.00 at
18 7 but that by for if from ! 93.33 for
19 6 but that by if from ! 93.33 !
20 5 but that by if from 93.33 if
21 4 but that by from 95.00 from
22 3 but that by 90.00 by
23 2 but that 78.33 that
24 1 but 50.00 but
Columns present parameters: (a) elimination stage, (b) number of characteristic features left, (c) set
of currently considered variables, (d) median predictive accuracy of the classifier (%), (e) attribute
selected to be eliminated
the distributions and dispersions of specific classification accuracies, we can assign
higher priority to these networks that have especially good results such as for example
100% recognition.
As before for forward selection, for backward elimination the search for subsets
of features is not exhaustive. Commencing with the entire set of variables we reject
one variable at a time with the decision based on the local context, and once some
variable is reduced it is not taken under consideration for the second time.
For the set of attributes with cardinality of 25 with arbitrarily assigned preference
orders, the decision algorithm generated within the approach of finding only minimal
cover performs rather poorly, correctly recognising barely a half of the testing sam-
ples. We can try to increase this accuracy by adjusting preference orders, yet there
are no quick procedures that could be employed to this end. When we induce all rules