example, if we were betting on the outcome of a game, and
the BN predicted Win 45% Draw 45% Loss 10% then this
would indicate a likely win for an each way bet. However,
such an analysis of the potential value of a shared highest
probability prediction is beyond the scope of this paper.
We divided the match data into disjoint subsets so that
some could be used for training and separate data used to
check the accuracy of the learners. The data for each season
was divided up into three groups of ten matches and one
group of eight matches, organised chronologically. We
maintain the ordering of games and always organise the
training so that the training data set are chronologically
immediately before the test data set. For comparison we
also used each complete season’s data for training and test
set for the learners. This again prejudices the results against
the expert BN because this will tend to overestimate the
accuracy of all the other learners. The machine learners
were tested with both our general model data and with the
data used by the expert BN. Using the two data sets allows
for a direct comparison with the same, expert chosen, data
set and a more general comparison with a data set a non
expert might choose. The results for both the general data
and the expert chosen data, shown in Tables 1 and 2, are
similar. Where changes in classiWcation error are mentioned
they are relative to the error obtained by choosing the most
common result from the training data.
example, if we were betting on the outcome of a game, and
the BN predicted Win 45% Draw 45% Loss 10% then this
would indicate a likely win for an each way bet. However,
such an analysis of the potential value of a shared highest
probability prediction is beyond the scope of this paper.
We divided the match data into disjoint subsets so that
some could be used for training and separate data used to
check the accuracy of the learners. The data for each season
was divided up into three groups of ten matches and one
group of eight matches, organised chronologically. We
maintain the ordering of games and always organise the
training so that the training data set are chronologically
immediately before the test data set. For comparison we
also used each complete season’s data for training and test
set for the learners. This again prejudices the results against
the expert BN because this will tend to overestimate the
accuracy of all the other learners. The machine learners
were tested with both our general model data and with the
data used by the expert BN. Using the two data sets allows
for a direct comparison with the same, expert chosen, data
set and a more general comparison with a data set a non
expert might choose. The results for both the general data
and the expert chosen data, shown in Tables 1 and 2, are
similar. Where changes in classiWcation error are mentioned
they are relative to the error obtained by choosing the most
common result from the training data.
การแปล กรุณารอสักครู่..