4.2.1. Complete seasons
For the 1995/1996 season the Naive Bayesian learner
correctly predicted the result of 26 and 22 of the 38 games
in the general and expert models respectively. This is a
reduction in the classiWcation error of about 26.31% and
15.78%. The naive Bayesian classiWer gives no direct indication
of the importance of any given attribute. However,
looking at the NPT for the classiWer in the general model
we can see that the six most signiWcant attributes in
descending order are: Team Ranking, Dozzell, Edinburgh,
Anderton, Dumitrescu, and Calderwood. There is some,
limited, agreement between MC4 and the naive Bayesian
learner on the signiWcant attributes, they agree on the two
most important of the thirty attributes for the 1995/1996
season. For the 1996/1997 season the Naive Bayesian
learner correctly predicted the result of 31 and 25 of the 38
games for the general and expert models, respectively. This
is a reduction in the classiWcation error of about 34.21%
and 18.42%.