A Goal-Based Approach for Learning in Business Processes - Intentional Perspectives on Information Systems Engineering - page 241

definition , namely, groups of instances whose contextual property-based similarity
can predict some behavioral similarity.
Behavioral similarity of process instances can be assessed using some path and
state similarity measures. Consequently, process instances can be grouped into clus-
ters of behaviorally-similar process instances, sharing similar paths and similar
termination states (outcomes). Each process instance in the experience base would
belong to one behavioral similarity cluster.
Contextual property-based similarity of process instances is possible when at
least one contextual property of these instances has similar values. The possible
number of contextual property-based similarity groupings is combinatorial in the
number of contextual properties. Not all these groupings are meaningful in terms of
behavior (e.g., grouping process instances based on the color of the customer's eyes
would probably be ineffective for predicting behavior of process instances).
Based on these two types of similarity, we define a context group as a group of
process instances, which are contextual property-based similar, and for which taking
similar paths implies achieving similar outcomes.
Note that this definition relates to a situation where the behavior of process
instances is fully consistent with respect to their context, namely, there are no unpre-
dicted behaviors or noisy data. Clearly, this is not the situation in real life, where
there might be “hidden” variables which cannot be tracked (e.g., distractions of
the machine operator) that affect the outcomes of the process. Hence, we cannot
assume full predictability of the outcomes given a context group and a process path.
However, we may assume that contextual properties have a certain effect and can
explain at least part of the variance in the outcomes achieved. Hence, for practical
purposes we can formulate the following postulate:
Postulate 1 : Consider two groups of process instances, PI 1 and PI 2 , so each
group includes contextual property-based similar process instances. Now consider
C 1 ⊂

PI 2 , so the paths taken by instances in C 1 and C 2 are all similar.
If statistical tests show that the termination states of C 1 and C 2 are not of the same
population, then PI 1 and PI 2 are in different context groups.
Postulate 1 gives us a criterion for excluding two sets of instances from being in
the same context group. It can be applied to groups of instances that follow similar
paths. However, we may have groups that follow different paths. In that case, we
assume the choice of path reflects some implicit domain knowledge used by the
process operators. This is reflected in the following postulate.
Postulate 2 : Groups of contextually similar process instances form one con-
text group if the distribution of their behavioral similarity categories is similar (not
significantly different).
The two postulates can be helpful when some grouping based on contextual
properties is available. However, as discussed above, the number of such group-
ings is combinatorial in the number of known contextual properties. To overcome
this difficulty, we employ a learning algorithm, which grows a decision tree whose
independent variable is the contextual properties while the dependent variable is
the behavioral similarity category of process instances. The algorithm is applied
through the following procedure [ 8] .

PI 1 and C 2 ⊂