Information Technology Reference
In-Depth Information
Table 6.5.
Overview of constructed learning algorithms by CAMLET to the datasets
of the rule sets learned from the UCI benchmark datasets
Distribution I
Distribution II
Distribution III
original
classifier set
overall
control structure
final
eval. method
original
classifier set
overall
control structure
final
eval. method
original
classifier set
overall
control structure
final
eval. method
Win+Boost+CS
Weighted
Voting
Weighted
Voting
Weighted
Voting
anneal
C4.5 tree
C4.5 tree
Boost+CS
C4.5 tree
Boost+CS
Weighted
Voting
audiology
ID3 tree
Boost
Voting
Random Rules CS+GA
Random Rules Simple Iteration
Best Select.
Weighted
Voting
Weighted
Voting
Weighted
Voting
autos
Random Rules Win+Iteration
ID3 tree
Boost+Iteration
Random Rules Boost
balance-
scale
Weighted
Voting
Random Rules Boost
Voting
Random Rules Boost+CS
Random Rules CS+GA
Voting
breast-
cancer
Boost+CS
+Iteration
Weighted
Voting
Weighted
Voting
Random Rules GA+Iteration
Voting
ID3 tree
Random Rules Win+Iteration
Weighted
Voting
Weighted
Voting
breast-w
ID3 tree
Win
ID3 tree
Iteration
Best Select. ID3 tree
CS+Iteration
colic
Random Rules CS+Win
Voting
ID3 tree
Win+Iteration
Best Select. ID3 tree
Win+Iteration
Voting
credit-a C4.5 tree
Win+Iteration
Voting
Random Rules Win+Iteration
Best Select. ID3 tree
CS+Boost+IterationBest Select.
CS means including reinfoecement of classifier set from Classifiser Systems
Win means including methods and control structure from Window Strat
Boost means including methods and control structure from Boosting
GA means including reinforcement of classifier set with Genetic Algorit
distributions. The class distribution for “Distribution I” is
P
=(0
.
35
,
0
.
3
,
0
.
3)
where
p
i
is the probability of class
i
. Thus, the number of class
i
instances in
each dataset
D
j
become
p
i
D
j
. Similarly, the probability vector of “Distribution
II” is
P
=(0
.
3
,
0
.
5
,
0
.
2) and that of “Distribution III” is
P
=(0
.
3
,
0
.
65
,
0
.
05).
Constructing proper learning algorithms for rule sets from UCI
datasets.
In the same way as the construction of an appropriate learning algo-
rithm for the meningitis data mining result, we constructed appropriate learning
algorithms for the datasets of rule sets from the eight UCI datasets. Table6.5
shows an overview of the constructed learning algorithms for each dataset, which
had three different class distributions.
For these datasets, CAMLET constructed various learning algorithms based
on 'random rule set generation', ID3 decision tree, and C4.5 decision tree. There-
fore, these learning algorithms consisted of new combinations of methods that
had previously never been seen in learning algorithms. Most of the learning algo-
rithms include 'Voting' from bagging or 'Weighted Voting' from boosting. With
regard to these results, CAMLET constructed selective meta-learning algorithms
for the datasets with the three different class distributions.
Accuracy Comparison on Classification Performances.
For the above
mentioned datasets, we used the five learning algorithms to estimate whether
their classification results reached or exceeded the accuracies when just pre-
dicting each majority class. Table 6.6 shows the accuracies of the nine learning
algorithms applied to each class distribution of the three datasets. The learn-
ing algorithms constructed by CAMLET, boosted J4.8, bagged J4.8, J4.8, and
BPNN always performed better than just predicting the majority class of each
dataset. In particular, Bagged J4.8 and Boosted J4.8 outperformed J4.8 and
BPNN for almost all datasets. However, their performances were suffered from
probabilistic class distributions for larger datasets, such as balance-scale and
credit-a.
