Information Technology Reference
In-Depth Information
be better classifier for the entire dataset than these hyper-plane learners and
bagged J4.8, they needed more training instances to become accurate classi-
fiers. Looking at the result of learning algorithm constructed by CAMLET, this
algorithm achieves almost the same performance as bagged J4.8, with smaller
training subset. However, it can outperform bagged J4.8 with larger training
subsets. Although the constructed algorithm was based on boosting, the com-
bination of a reinforcement method from Classifier Systems and the outer loop
was able to overcome the disadvantage of boosting for a smaller training subset.
Rule evaluation models for the meningitis data mining result dataset.
In this section, we present rule evaluation models for the entire dataset learned
using CAMLET, OneR, J4.8 and CLR. This is because they are represented as
explicit models such as a rule set, a decision tree, and a linear model set.
As shown in Fig. 6.5, the indices used in the learned rule evaluation models
are taken, not only from a group of indices that increases with the correctness of
a rule, but also from different groups of indices. Indices such as YLI1, Laplace
Correction, Accuracy, Precision, Recall, Coverage, PSI and Gini Gain are indices
that were formerly used for models. Later indices include GBI and Peculiarity,
which sums up the difference in antecedents between one rule and the other rules
in the same rule set. This corresponds to a comment made by the human expert.
He said that he evaluated these rules not only according to their correctness but
also their interestingness based on his expertise
Top 10 frequency in OneR models
Top 10 frequency in CAMLET models
Peculiarity
Peculiarity
RelativeRisk
GBI
ChiSquare-one
MutualInformation
AddedValue
GOI
GBI
OddsRatio
Accuracy
Precision
Lift
LaplaceCorrection
BI
GiniGain
Coverage
KI
BC
YLI1
0
1000000 2000000 3000000 4000000 5000000
0
200
400
600
800
1000
1200
1400
1600
Top 10 frequency in CLR models
Top 10 frequency in J4.8 models
Peculiarity
Peculiarity
GBI
LaplaceCorrection
MutualInformation
OddsRatio
J-Measure
Prevalence
Coverage
RelativeRisk
Precision
Precision
GBI
Recall
Accuracy
Credibility
LaplaceCorrection
GiniGain
Lift
OddsRatio
0
10000
20000
30000
40000
50000
0
5000
10000
15000
20000
Fig. 6.5.
Top 10 frequencies of the indices used by the models of each learning algo-
rithm with 10000 bootstrap samples of the meningitis datamining result dataset and
executions
