Supporting Dynamic, People-Driven Processes through Self-learning of Message Flows - Advanced Information Systems Engineering - page 656

emails. When organizing a large event, these messages are the main artifacts
to coordinate between participants. While the necessary activities are relatively
clear, the type of messages, their occurrence, and evolution will remain dynamic.
We address three major problems (i) learning of message-activity dependencies,
(ii) prediction of which messages will arrive, and (iii) automatically reflecting
work evolution in the process model. The subsequent challenges are then to
distinguish between accidental one-time deviations and desired process improve-
ments, managing the interleaving of messages, and the inability to observe the
complete set of user actions.
Our contributions in this paper are (i) a probabilistic process model and exe-
cution engine that does not rely on a precise occurrence or absence of messages;
(ii) a self-learning (thus unsupervised) message flow algorithm to detect message-
activity dependencies and updates thereof; and (iii) a message recommendation
mechanism to support the analysis of unstructured and semi-structured mes-
sages. The main applied approaches are log sequence mining, and providing —
respectively extracting — message-activity correlation information.
Our contributions bring benefit to both process designers and process users.
Process designers need not capture all possible deviations nor the exact mapping
of input and output messages to activities as this is automatically learned from
process users. These in turn see the immediate effect of their applied expertise
in form of process model changes without having to include a dedicated process
designer.
The remainder of this paper is structured as follows: A motivating scenario
sets the scene for our self-learning message flow algorithm (Section 1.1). We
discuss related work in Section 2. Section 3 introduces our approach, followed by
the probabilistic models in Section 4. Section 5 describes the recommendation
and learning mechanism, which we evaluate in Section 6. We provide a short
conclusion and outlook on future work in Section 7.
1.1 Motivating Scenario
The example in Figure 1 depicts a flexible people-driven order process. The
individual work steps describe a general order of user activities to successfully
complete a process instance. The outlined flow, however, does not enforce the
exact order of activities, which is up to the user, and covers no exceptions or
process adaptations that might arise due to specific customer request, incomplete
information, or user specific expertise. Consequently, the listed document types
that characterize the exchanged messages specify merely an initial set of expected
documents. The process visualization in Figure 1 does not apply any particular
process modeling language but rather presents an intuitive view on the involved
documents that represent input and output of activities. In this scenario, we
encounter various forms of message flow evolution:
Missing Documents : When the Replenish step (C) is updated to make use
of an automatic restocking system, only user confirmation is required and
the Quote message (3) no longer occurs.