Thoughts on:

      "Mining Association Rules from Set-Valued Data",
      Christopher A. Shoemaker,
      Department of Computer Science, 
      Worcester Polytechnic Institute, 2001

I do not have an electronic copy of this paper.

This thesis addresses the idea of having set-valued attributes in 
addition to single valued attributes from which to mine association
rules. This idea is similar to our own data but here set-valued
attributes consist of an unordered list of values when combined, 
form a single value for that attribute. No values are repeated in this
unordered list.
      
Shoemaker extends ARMiner's implementation of Apriori to handle
set-valued attributes. We might take this and extend it to handle
our time-series attributes. If not, it will still serve well as an
example of working with and extending ARMiner.

The main differences between set-valued attributes and our time
series attributes are: sequences are ordered in time and this order
has meaning, our time series attribute values are expected to be
mostly numeric and we wish to treat those numeric values separately
at times, depending on the specific approach rather than as a single
attribute value.

It looks like we may pre-process the time-series attributes for now.
If so, ARMiner could be used without modification for our initial
implementation.