Clustering the incident space: While the previous model
is appealing by its simplicity, it completely ignores the distribution of the aggregated feature vector, in spite of it giving
precious information about the possible states of the monitored
system, which should intuitively have a role in the decision.
While it is intuitive that a decision may depend on some kind
of state, it is too coarse to simply identify decision (class) and
state, as Naive Bayes does. What could possibly be identified
with the state is some structure capturing the hypothetical
reasoning which leads from the observation of the incidents
to the decision, but not the decision itself.