Information Technology Reference
In-Depth Information
Kolmogorov-Smirnov test using a window size of
w
= 400 is applied to the
vector corresponding to each activity pair in
Σ
Σ
. Fig. 7 depicts the aver-
age significance probability of
KS
-test on all activity pairs. It could be seen
that significant troughs are formed at indices 1200, 2400, 3600 and 4800. These
are indeed the points where the models have been changed. Thus the features
and approach proposed in this paper are shown to have significant promise in
accurately identifying the points of change.
×
Fig. 7.
Average significance probability (over all activity pairs) of
KS
-test on
J
-
measure estimated for each trace.
X
-axis represents the trace index.
Y
-axis represents
the significance probability of the test. Troughs signify change points.
The second objective in handling concept drift is that of
change localization
.
In order to localize the changes (identify the regions of change), we need to con-
sider each activity pair individually or a subset of activity pairs. For example,
the change from
M
1
to
M
2
is localized in the region pertaining to high insur-
ance claim checks. We expect characteristic changes in features pertaining to
these activities and other activities related to these activities. For example, in
M
1
, the activities 'High Medical History Check' and 'Contact Hospital' always
follow the activity 'Register' whenever a claim is classified as high. In contrast,
in
M
2
, these activities need not always follow 'Register' due to the fact that
both these activities are skipped if 'High Insurance Check' fails while 'Contact
Hospital' is skipped if 'High Medical History Check' fails. During simulation, we
have set the probability of success of a check to 90%. We have considered the
window count (WC) feature for the activity relation 'Contact Hospital' follows
'Register' on a window size of 10 in each trace separately. Fig. 8a depicts the
significance probability of the univariate
KS
-test using a window size of
w
= 200
on this feature. It could be seen that one dominant trough is formed at index
1200 indicating that there exists a change in the region between 'Register' and
'Contact Hospital'. No subsequent changes with respect to this activity pair is
noticed which is indeed the case in the models.
As another example, we have considered the activity 'Prepare Notification'
along with all the three 'Send Notification' activities. There exists a change
pertaining to these activities between models
M
2
and
M
3
,
M
3
and
M
4
,and
M
4
and
M
5
. More specifically, we have considered the window count feature on
the activity relations 'Send Notification By Phone' follows 'Prepare Notification',
'Send Notification By email' follows 'Prepare Notification' and 'Send Notification
