Weighting of Features by Sequential Selection - Feature Selection for Data and Pattern Recognition - page 83

Table 5.3 Backward elimination of attributes basing on the performance of reduced all rules on
examples decision algorithm
(a)

(b) (c) (d) (e) (f) (g) (h)

0 25 but and not in with on at of as this that by for to if 46 , 191 41 80 76.67 and

what from . , ; : ! ? ( -
1 24 but not in with on at of as this that by for to if what 42 , 018 33 19 83.33 (

from . , ; : ! ? ( -
2 23 but not in with on at of as this that by for to if what

34 , 390 30 38 85.00 on

from . , ; : ! ? -
3 22 but not in with at of as this that by for to if what from 24 , 732 29

35 86.67 as

.,;:!?-
4 21 but not in with at of this that by for to if what from . ,

19 , 869 29 33 86.67 this

;:!?-
5 20 but not in with at of that by for to if what from . , ; : ! 15 , 072 29 32 86.67 -

?-
6 19 but not in with at of that by for to if what from . , ; : ! ? 12 , 407 29 31 86.67 .

7 18 but not in with at of that by for to if what from , ; : ! ? 9 , 065 29 31 86.67 ;

,

8 17 but not in with at of that by for to if what from , : ! ? 6 597 29 31 86.67 with

9 16 but not in at of that by for to if what from , : ! ? 4 , 939 29 31 86.67 :

10 15 but not in at of that by for to if what from , ! ? 3 , 740 29 31 86.67 for

11 14 but not in at of that by to if what from , ! ? 3 , 031 29 31 86.67 if

12 13 but not in at of that by to what from , ! ? 2 , 456 29 31 86.67 what

13 12 but not in at of that by to from , ! ? 2 , 131 29 31 86.67 that

14 11 but not in at of by to from , ! ? 1 , 841 29 31 86.67 but

659 29 31 86.67 in
16 9 not at of by to from , ! ? 1 , 085 29 21 85.00 ?
17 8 not at of by to from , ! 861 21 55 88.33 at
18 7 not of by to from , ! 649 21 54 88.33 !
19 6 not of by to from , 407 16 88 88.33 to
20 5 not of by from , 311 13 106 90.00 ,
21 4 not of by from 172 13 84 83.33 of
22 3 not by from 100 11 71 76.67 not
23 2 by from 34 16 22 63.33 by
24 1 from 4 8 4 23.33 from
Columns present parameters: (a) elimination stage, (b) number of characteristic features left, (c) set
of currently considered variables, (d) number of rules in a decision algorithmwithout any constraints,
(e) minimal support required of DRSA rules resulting inmaximal classification accuracy, (f) number
of exact DRSA rulesmeeting constraints on support, (g)maximal predictive accuracy of the classifier
(%), (h) attribute selected to be eliminated

15 10 not in at of by to from , ! ? 1 ,

When we compare predictive accuracies of rule classifiers tested at each reduction
stage, plotted in Fig. 5.4 , not by exact numbers but perceivable trends, to those pre-
viously studied in forward selection approach, it is immediately apparent that they