Information Technology Reference
In-Depth Information
Table 6.2.
Description of the meningitis datasets and the results of data mining
#Mined
Dataset
#Attributes #Class
rules #'I' #'NI' #'NU'
Diag
29
6
53
15
38
0
C Course
40
12
22
3 8
1
Culture+diag
31
12
57
7 8
2
Diag2
29
2
35
8 7
0
Course
40
2
53
12
38
3
Cult find
29
2
24
3 8
3
TOTAL
—
—
244
48
187
9
datasets, some appearances of meningitis patients were considered to be at-
tributes and the diagnosis of each patient was considered as a class. Each rule
set had been mined using appropriate rule induction algorithms composed by a
constructive meta-learning system called CAMLET [12]. We labeled each rule
with one of three evaluations (I: Interesting, NI: Not-Interesting, NU: Not-
Understandable) based on evaluation comments provided by a medical expert.
Constructing a proper learning algorithm to construct the menin-
gitis rule evaluation model.
We developed a constructive meta-learning
system called CAMLET [11] to choose an appropriate learning algorithm for a
given dataset using a machine learning method repository. To implement the
method repository, we first identified each functional part, called method, from
the following eight learning algorithms: Version Space [31], AQ15 [32], Classifier
Systems [33], Neural Network, ID3 [34], C4.5, Bagging and Boosting. With the
method repository, CAMLET constructs an appropriate learning algorithm for
a given dataset by searching through the possible learning algorithm specifica-
tion space obtained by the method repository for the best one, using a Genetic
Algorithm.
After the initial population was set up
τ
= 4 and a number of refinement were
made
N
= 100, CAMLET searched through up to 400 learning algorithms, from
6000 possible learning algorithms, for the best one. Fig.6.3 shows the algorithm
constructed by CAMLET for the dataset of the meningitis data mining result.
This algorithm iterates boosting of a C4.5 decision tree for randomly split
training datasets. Each classifier set generated by the C4.5 decision tree learner
is reinforced with a method from Classifier Systems. Then the learned committee
aggregates with weighted voting from boosting.
Comparison of the classification performances.
In this section, we present
the results of accuracy comparison over the entire dataset, the recall of each class
label, and their precisions. Since Leave-One-Out uses just one test instance and
the remainders are used repeatedly as the training dataset for each instance of
a given dataset, we could evaluate the performance of a learning algorithm for
a new dataset without any ambiguity.
