A Goal-Based Approach for Learning in Business Processes - Intentional Perspectives on Information Systems Engineering - page 243

Considering our bottle manufacturing running example, one of the difficulties
faced is the large number of possible process paths (considering each one of the
14 machines and 40 employees who operate the machines as different paths).
Furthermore, the selection of a machine and a worker at runtime is mainly done
based on availability, and to a lesser extent on the context, so this choice is not
expected to reflect the relevant contextual properties we are looking for. Still, this
choice might affect the outcomes. To overcome this, we decided to identify a sub-
set of very reliable employees and use only process instances they participate in
for context identification. We also decided not to differentiate paths where different
machines were used, but to include the time since the last maintenance operation of
the machine as a contextual property.
The identification procedure is applied as follows:
Step 1 : An initial classification of the process instances related to whether or
not the process faced an event of problem identification. Clearly, this is a
contextual variable that affects the process behavior. Hence, we separately
performed the following steps to instances where problems were identified
and to instances where production was performed without interrupts. We
demonstrate the next steps with respect to the group where no problems
occurred.
Step 2 : Paths were clustered (disregarding machines and workers, as discussed
above). The termination states were divided into two groups: (1) instances
where the customer accepted the delivery without a need for a 100% inspec-
tion, (2) instances where the customer accepted the delivery after a 100%
inspection, or where the customer rejected the delivery or where the delivery
was cancelled. The combination of path similarity groups and termination
state groups included 12 behavioral categories.
Step 3 : Applying the decision tree growing algorithm resulted in the tree
depicted in Fig. 3.
Each path in the tree (from the root to a leaf) represents a combination
of contextual properties relevant for the behavior of process instances. Each
node in the tree holds a set of process instances that can be characterized
by a distribution over behavioral similarity categories. For example, node 13
stands for process instances related to products in the food and cosmetics
market whose size is large and where the customer required resistance of the
bottle to high temperatures. Node 12 represents process instances in the med-
ical supplies market with special covers (children proof) where the machine
used was not within a short period after its periodic maintenance (hence its
maintenance state is not considered as best).
Step 4 : Applying postulates 1 and 2. Due to space limitation, we only demon-
strate Step 4 with respect to parts of the tree, leaf nodes 8, 9, and 13.
The behavioral categories of the instances in all three nodes fall into three
path similarity categories (paths 1-3) and two termination state categories,
distributed as shown in Table 2.