4.2.2. Separate training and test data – single season
The results for the 1995/1996 season showed the average
classiWcation error to be 7.41% and 3.70% higher for the
general and expert data sets, respectively. However, for the
1996/1997 season the general model classiWcation error was
7.41% lower while that for expert data set model increased
by 3.70%. Most classiWers achieved better results for the
1996/1997 season than the 1995/1996 season which may
indicate greater stability in the team in the later season.
4.2.3. Separate training and test data – cross seasons
The cross season results for the naive Bayesian learner
were roughly comparable to its in-season results. Overall it
achieved a classiWcation accuracy of 33.09% and 35.29% for
the general and expert models which only bettered the most
common classiWer by 0.98% and 3.18%, respectively. Ignoring
the case using the same training and test data for the
complete seasons, the naive Bayesian learner came out second
best overall on the general model and Wfth overall on
the expert model.
4.3. Data driven Bayesian learner
The BNs for the data driven Bayesian learner were generated
using the structural learning wizard from the Hugin
Developer version 6.1 program. The process used was to
run the program using an initial Level of Significance
of 0.1. If no link directed to the result node was
formed the process was rerun doubling the Level of
Significance until a network with at least one link
directed to the result node was achieved. Since, in this problem
all of the nodes except the result node have their values
speciWed any nodes in the network with no links directed to
the result node were removed. The remaining network was
used for the testing. The overall classiWcation error of the
various learnt networks for disjoint training and test data
sets was 67.69% and 67.38% for the general and expert
models, respectively.