A Goal-Based Approach for Learning in Business Processes - Intentional Perspectives on Information Systems Engineering - page 240

improvement recommendations to the process model. Note that these improvements
are not automatically deployed. Rather, the analysis yields improvement recommen-
dations which should first be reviewed by humans and only then used for updating
the process model.
The runtime phase is an ongoing phase of process execution. Every new pro-
cess instance is classified into a context group and follows the path recommended
accordingly. Its context, path, and outcome data are then stored in the experience
base. There might be instances which cannot be classified into an existing context
group; they are executed and their data is also stored in the experience base. In
some cases, specifically when facing unexpected external events, the process oper-
ators may decide to deviate from the process model and take a path which has not
been taken before. These are also stored in the experience base. Periodically, when
a considerable number of new experiences have been added to the experience base,
learning can be applied again, triggering a new cycle. In what follows we provide
details about the phases of the learning cycle.
4.1 Context Identification
As explained above, the challenge in identifying context is the huge amount of
contextual information that may be available. We seek for classification criteria of
process instances that would be effective in determining the best process path at a
given situation. This classification should also be meaningful in business terms, so
each group of instances can be characterized based on its contextual properties.
Recall, the data of the process instances stored in the experience base includes
their actual path, their outcome (or termination state), and their contextual informa-
tion. The path and the termination state of a process instance constitute its behavior.
In a perfect world, process instances that have similar contexts would follow similar
paths to lead to a given termination state. However, our knowledge of the process
(relevant) context is partial. Under partial knowledge, we may not be aware of con-
textual variables whose different values may differently affect the process behavior,
and can be considered “different contexts”. Lacking such knowledge, we may group
process instances that partially share the same context but exhibit different behav-
iors. This would not be an effective strategy for learning the best paths that for a
given context would achieve desirable outcomes. Hence, with the knowledge that
exists at this phase, process instances can be grouped considering two types of
similarities:
(1) Behavioral similarity.
(2) Contextual property-based similarity.
Clearly, these two groupings are expected to be different, since not all contex-
tual properties necessarily affect process behavior, and some properties may have
a similar effect. Our interest is to identify a third type of grouping, context groups