Information Technology Reference
In-Depth Information
change
points
M
1
M
2
M
3
M
4
M
5
1
1200
2400
3600
4800
6000
Fig. 5.
Event log with traces from each of the five models juxtaposed. Also indicated
are change points between models.
are observed around indices 24, 72 and 96 which are indeed the points of change
(remember that we have split the log into 120 sub-logs with the change points at
indices 24, 48, 72 and 96). The change at index 48 corresponding to the transition
from
M
2
to
M
3
could not be uncovered using this feature set due to the fact
that the relation type counts would be alike for logs generated from these two
process variants.
a
b
Fig. 6.
(a) Significance probability of Hotelling
T
2
test on relation counts (b) Average
significance probability (over all activity pairs) of
KS
-test on
J
-measure. The event
log is split into sub-logs of 50 traces each.
X
-axis represents the sub-log index.
Y
-axis
represents the significance probability of the test. Troughs signify change points.
We have computed the
J
-measure for each sub-log and for every pair of ac-
tivities,
a
,
b
in
Σ
(
a
F
b
,
b
follows
a
within a window of size 10). The univariate
Kolmogorov-Smirnov test using a window size of
w
=10isappliedonthe
J
-
measure of each activity pair. Fig. 6b depicts the average significance probabil-
ity of
KS
-test on all activity pairs. It could be seen that significant troughs are
formed at indices 24, 48, 72 and 96 which correspond to the actual change points.
Unlike the relation type count feature, the
J
-measure feature is able to capture
all the four changes in the models. This can be attributed to the fact that the
J
-measure uses the probability of occurrence of activities and their relations. In
M
2
, there could be cases where all the modes of notification are skipped (XOR
construct). However in
M
3
at least one of the modes need to be executed (OR
construct). This results in a difference in the distribution of activity probabilities
and their relationship probabilities which is elegantly captured in the
J
-measure.
We have considered the
J
-measure for each trace separately instead of at
the sub-log level. Each activity pair generates a vector of dimension 6000 cor-
responding to the
J
-measure of that activity pair in each trace. The univariate
