I have started the process of building my association rules model
(based on the Apriori method, in R). I have primarily relied on lift as
my metric in this case, but it is a little bit tricky, because the range
for lift is 0 to infinity, with the value of 1 meaning the sides are
perfectly independent of each other. There's not much room on the
left side of that equation, but infinite room on the right. The idea of
lift is that the higher the lift value of a particular rule, the more
significant is the positive relationship between the left and right
sides of the rule. My lifts were mostly between 1 and 1.5--not very
high so far. Moreover, the most significant rules highlight the same
small subset of variables. I've used methods to drop redundant rules
and to raise/lower the threshold for serving up a rule, but this kind
of tuning has been of limited value. I am going to try to do much
narrower slices of the data to see if I can find more significant rules
for prediction. I suspect that I currently have too many factors--and
therefore too much noise--in my model. My next exploration will be
in conditional inference trees that are able to on their own drop less
significant factors and create non-overlapping classes.
