Page 255 -

8.6 Bringing It All Together 237

information as a predictor variable. Contextual information is information that
is not in the event log and that is not necessarily related to a particular case.
For example, the weather may influence a decision. This can only be discov-
ered if the weather condition is taken into account as a predictor variable.
Decisions may also depend on the volume of work in the pipeline. One can
imagine that the choice between b and c in Fig. 8.9 depends on the work-
load of the two experts Sue and Sean. When they are overloaded, it may be
less likely that b is selected. These examples illustrate that predictor variables
are not limited to case and event attributes. However, note the “curse of di-
mensionality” discussed in Sect. 3.6.3. Analyzing decision points with many
predictor variables may be computationally intractable.
In Figs. 8.14 and 8.15, we used classification to learn decision rules. The pre-
dictor variables can also be used to learn other properties of the process.For
instance, one may be interested in characterizing cases for which a particu-
lar activity is executed. Classification can also be used to uncover reasons for
non-conformance. As shown in Fig. 7.8, the event log can be split into two
sublogs: one event log containing only fitting cases and one event log con-
taining only non-fitting cases. The observation whether a case fits or not, can
be seen as a response variable. Hence, classification techniques like decision
tree learning can be used to characterize cases that deviate. For example, one
could learn the rule that cases of gold customers from the southern region tend
to deviate from the normative model. Similarly, one could learn rules related
to the lateness of cases. For instance, one could find out that cases involving
Ellen tend to be delayed.
These examples show that established classification techniques can be com-
bined with process mining once the process model and the event log are con-
nected through replay techniques.

8.6 Bringing It All Together

In this chapter, we showed that a control-flow model can be extended with additional
perspectives extracted from the event log. Figure 8.16 sketches the approach to ob-
tain a fully integrated model covering all relevant aspects of the process at hand.
The approach consists of five steps. For each step, we provide pointers to chapters
and sections in this book:
• Step 1: obtain an event log. Chapter 4 showed how to extract event data from a
variety of systems. As explained using Fig. 4.1, this is an iterative process. The
dotted chart described in Sect. 8.2 helps to explore the event log and guide the
filtering process.
• Step 2: create or discover a process model. Chapters 5 and 6 focus on techniques
for process discovery. Techniques such as heuristic mining and genetic mining