Handling Concept Drift in Process Mining - Advanced Information Systems Engineering - page 401

a b

Fig. 8. (a) Significance probability of KS -test on WC feature estimated for the relation,
'Contact Hospital' follows 'Register'. Trough indicate change point w.r.t this feature.
(b) Average significance probability (over activity pairs) of KS -test on WC feature
estimated for the various modes of 'Send Notification' follows 'Prepare Notification'
relation. Troughs indicate change point w.r.t these activities. X -axisrepresentsthe
trace index. Y -axis represents the significance probability of the test.

By Post' follows 'Prepare Notification'. Fig. 8b depicts the average significance
probability of the univariate KS -tests using a window size of w = 200 on the WC
feature of these three activity pairs. We see three dominant troughs at around
indices 2400, 3600 and 4800 signifying the changes in the models. Certain false
alarms (minor troughs) can also be noticed in this plot. One means of alleviating
this is to consider only those alarms with a significance probability less than a
threshold, δ . In this fashion, by considering activities (and/or activity pairs) of
interest, one can localize the regions of change. Furthermore, one can also get
answers to diagnostic questions such as “ Is there a change with respect to activity
a in the process at time period t ”?

6Ou ook

Dealing with concept drifts raises a number of scientific and practical challenges.
In this section, we highlight some of these challenges.

- Change-Pattern Specific Features: In this paper, we presented very generic
features (based on follows/precedes relation). These features are neither com-
plete nor sucient to detect all classes of changes. An important direction of
research would be to define features catering to different classes of changes
and investigate their effectiveness. A taxonomy/classification of change pat-
terns and the appropriate features for detecting changes with respect to those
patterns is needed. For example, if we would like to detect changes pertain-
ing to a loop construct (insertion/removal/modification of loops as changes
in process variants), tandem arrays [16] would be an appropriate feature to
consider.
- Feature Selection: The feature sets presented in this paper result in a large
number of features. For example, the activity relation count feature type
generates 3

| Σ |

features whereas the window count and J -measure generate

2 features (corresponding to all activity pairs). On the one hand, such high

| Σ |