Page 141 -

130 M. Adams
relevant influencing factors and relationships between the inner state and the exter-
nal environment.
A taxonomy of contextual data that may be recorded and applied to a work-
flow instance may be categorized as follows (examples are drawn from the Order
Fulfillment process):

Generic (case independent): Data attributes that can be considered likely to
occur within any process (of course, the data values change from case to case).
Such data would include descriptors such as when created, created by, times
invoked, last invoked, current status; and role or agent descriptors such as expe-
rience, skills, rank, history with this process and/or task, and so on. Process
execution states and process log data also belong to this category.
Case dependent with a priori knowledge: The set of data that are known to
be pertinent to a particular case when it is instantiated. Generally, this data set
reflects the data variables of a particular process instance. Examples are customer
name, address, and delivery location; freight costs, size, and weight; ordered item
names, descriptions, costs, etc.; and deadlines both approaching and expired.
Case dependent with no a priori knowledge: The set of data that only becomes
known when the case is active and deviations from the known process occur.
Examples in this category may include complications that arise for incorrect
payments; unavailable stock, routes, and/or couriers; natural disasters preventing
delivery; and so on.

Methods for capturing contextual data typically focus on collecting a complete
set of knowledge from an “expert” and representing it in a computationally suitable
way. Such approaches depend heavily on the expert’s ability to interpret their own
expertise and express it in nonabstract forms. However, experts often have difficulty
providing information on how they reach a specific judgment, and will offer a jus-
tification instead of an explanation. Furthermore, the justification given varies with
the context in which the judgement was made.
Theories of context generally fall into two distinct groups: divide-and-conquer,
a top-down approach that views context as a way of partitioning a global model
into simpler pieces (e.g., the “expert” approach described above), and compose-and-
conquer, a bottom-up approach that holds that there is no tangible global model to
begin with, but only local perspectives, and so views context in terms of locality in
a (possible or potential) network of relations with other local perspectives.
A top-down, complete representation of knowledge within a given domain is con-
sidered by many researchers to be impossible to achieve in practice, and is perhaps
not even desirable. In terms of using context as a factor in computational decision
making, it is considered more judicious to capture only that subset of the complete
contextual state of a particular domain relevant to making a correct and informed
decision.
One bottom-up approach to the capture of contextual data that offers an alterna-
tive method to global knowledge construction is Ripple Down Rules (RDR), which
comprise a hierarchical set of rules with associated exceptions.