Information Technology Reference
In-Depth Information
Table 4.4.
The confusion matrix of the
Nursery
database using
PAT, C4.5 and OAT
a b c d e ¡- classified as
60088a=n m
00000b= mmend
20130c= ry m
200 43d=p y
4008 7e=sp -p r
a b c d e ¡- classified as
60088a=n m
00000b= mmend
20130c= ry m
200 43d=p y
200 0 7e=sp -p r
a b c d e ¡- classified as
401 25a=n m
00000b= mmend
00600c= ry m
003 60d=p y
000 4 5e=sp -p r
Table 4.5.
Tests performed on
lymphography
database
Threshold well classif.
¬
well classif. 50%
0.06
91.93%
8.06%
PAT 0.07
95.16%
4.83%
0.08
83.87%
12.90%
C4.5
79.03%
19.35%
01.61%
OAT
79.03%
19.35%
01.61%
35% for the attribute
parents
, 37% for
has-nur
, 39% for
health
and 13% for
form
.
We remark that our results are closer to those given by C4.5 when the Threshold
is 0.02 or 0.03. Our results do not improve when decreasing the threshold because
all the attributes in this database are independent. The confusion matrix are
shown in Table 4.4.
We find that
PAT
and
C4.5
give nearly the same classification error when
the attributes are independent.
OAT
is better when the class is
very-recom
.We
have also tested our approach on the
lymphography
database [20]. The training
set has 148 objects and 18 discrete attributes. The class takes 4 values:
normal,
metastases, malign-lymph, fibrosis
. We also use a test data which has 62 objects.
The missing values rates in the test data are: 56% for the attribute
block-of-affere
,
30% for
lym-nodes-dimin
, 40% for
changes-in-node
, 12% for
early-uptake-in
, 13%
for
special-forms
, 11% for
changes-in-stru
, 17% for
defect-in-node
and 11% for
the attribute
lym-nodes-enlar
. Table 4.5 contains the tests performed on the
lymphography
database [20] using the
PAT
approach, C4.5 and
OAT
.
From Table 4.6, we find that generally the performance of our
PAT
approach is
closer to those given by
C4.5
and
OAT
, but it is better than themwhen the class is
malign-lymph
. We have also tested our approach on the
Mushroom
database [20].
The training data has 5644 instances and 22 discrete attributes. The class takes
