Information Technology Reference
In-Depth Information
two independent tasks, viz., check insurance and check medical history need
to be executed. For high claims, three tasks need to be executed viz., check
insurance, check medical history, and contact doctor/hospital for verification. If
one of the checks shows that the claim is not valid, then the claim is rejected;
otherwise, it is accepted. An insurance grant and acceptance decision letter is
prepared in cases where a claim is accepted while a rejection decision letter is
created for rejected claims. In both cases, a notification is sent to the claimant.
Three modes of notification are supported viz., by email, by telephone (fax) and
by postal mail. The case should be archived upon notifying the claimant. This
can be done with or without the response for the questionnaire. However, the
decision of ignoring the questionnaire can only be made after a notification is
sent. The case is closed upon completion of archiving task.
Fig. 4 depicts five variants of this process represented in YAWL [14] notation.
The dashed rectangles indicate regions where a change has been done in the pro-
cess model with respect to its previous variant. The changes can have various
reasons. For example, in Fig. 4(a), the different checks for high insurance claims
are modeled using a parallel construct. However, a claim could be rejected if
any one of the checks fail. In such cases, the time and resources spent on other
checks go waste. To optimize this process, the agency can decide to enforce an
order on these checks and proceed on checks only if the previous check results
are positive. In other words, the process is modified with a
knockout
strategy
adopted for high insurance checks as depicted in Fig. 4(b). As another exam-
ple, the OR-construct pertaining to the sending of notification to claimants in
Fig. 4(c) has been modified to an exclusive-or (XOR) construct in Fig. 4(d).
The organization could have taken a decision to reduce their workforce as a
cost-cutting measure. Due to availability of limited resources, they would like to
minimize the redundancy of sending the notification through different modes of
communication and restrict it to only one of the modes.
Let us denote these process variants as
M
1
,M
2
,M
3
,M
4
and
M
5
.Wehave
modeled each of these process variants in CPN tools [15] and simulated 1200
traces for each model. We created an event log
of 6000 traces by juxtaposing
each set of the 1200 traces. The event log contains 15 activities or event classes
(i.e.,
L
|
Σ
|
L
,
our first objective is to
detect the four change points pertaining to these five process variants
as depicted
in Fig. 5.
The ideas presented in this paper have been implemented as the
concept drift
plugin in ProM. We have considered global features (at sub-log level) and local
features (both at trace and sub-log level) for our analysis. To facilitate this,
we have split the log into 120 sub-logs using a split size of 50 traces. We have
computed the relation type count (RC) of all 15 activities thereby generating a
multi-variate vector of 45 features for each sub-log. We have applied the Hotelling
T
2
hypothesis test on this multi-variate dataset using a moving window of size,
w
= 8. For this hypothesis test, we have randomly chosen 6 of the 45 features
with a 10-fold cross validation. Fig. 6a depicts the average significance probability
of the Hotelling
T
2
test for the 10 folds on this feature set. The troughs in the
plot signify that there is a change in the distribution of the feature values in
the log. In other words, they indicate that there is drift (change) in the concept,
which here corresponds to the process. It is interesting to see that the troughs
= 15) and 58953 events. Given this event log
