Using Cooperative Coevolution for Data Mining of Bayesian Networks - Advanced Techniques in Knowledge Discovery and Data Mining - page 170

6.5 Performance of CCGA

6.5.1 Experimental Methodology

A common practice to assess the performance of a Bayesian network learning
algorithm is to test the algorithm on data sets that are generated from known
network structures by probabilistic logic sampling [6.37]. We follow this prac-
tice and test our algorithm on seven different data sets. All of the data sets are
generated from well-known benchmark Bayesian networks, which include the
ALARM, the ASIA, and the PRINTD networks. Table 6.1 gives a summary
of the data sets that we used in our experiments.
Five of the data sets are generated from the ALARM network obtained
from different sources. Originally, the ALARM network was used in the med-
ical domain for potential anesthesia diagnosis in the operating room [6.38].
Because the network, with 37 nodes and 46 directed edges, has a complex
structure, it is widely used to evaluate the performance of a learning algo-
rithm. The PRINTD network is primarily constructed for troubleshooting
printer problems in the Windows TM operating system [6.3]. It has 26 nodes
and 26 edges. The ASIA-1000 data set is generated from the ASIA net-
work, which is a relatively simple structure that contains eight nodes and
eight edges. The network is also known as the “chest-clinic” network, which
describes a “fictitious medical example whether a patient has tuberculosis,
lung cancer, or bronchitis, related to their X-ray, dyspnea, visit-to-Asia, and
smoking status” [6.39], [6.40]. The data set contains 1000 cases.
In our experiment, we compare the performance of our algorithm with
MDLEP. All algorithms (including the implementation of MDLEP, obtained
from the authors) are implemented in the C++ language and compiled using
the same compiler. 10 ThesameMDLmetricevaluationroutineisusedsothat
the difference among implementations is minimized. For all algorithms, the
maximum size of a parent set is five. Because the algorithms are stochastic in
nature, they are executed 40 times for each test instance. All our experiments
are conducted on Sun Ultra-5 workstations.
10 Weusetheg++ compiler with -O2 optimization level.

Data set Original network Size MDL score of original

network

ALARM-1000 ALARM 1000 18,533.5

ALARM-2000 ALARM 2000 34,287.9

ALARM-5000 ALARM 5000 81,223.4

ALARM-10000 ALARM 10,000 158,497.0

ALARM-O ALARM 10,000 138,455.0

ASIA-1000 ASIA 1000 3416.9

PRINTD-5000 PRINTD 5000 106,541.6

Table 6.1. Data sets used in the experiments.