Information Technology Reference
In-Depth Information
On the one hand, our approach incorporates aspects of all three process types.
Our notion of people-driven process allows users to determine ad-hoc the process
execution order. There exists, however, a well structured process model as a
baseline guidance. Recommendations and learning is similar to semi-structured
systems as we apply past execution traces to dynamically update the process
model.
On the other hand, our approach deviates in several aspects from existing flex-
ible process recommendation systems. We relax the assumption of complete ob-
servability of user actions and instead rely on a combination of observed messages
and actions. The primary focus in this paper, however, lies on the self-adaptation
of message flows, leaving a reordering and adjustment of process activities aside.
We addressed this aspect of automatic people-driven process adaptation in our
previous work [5]. In addition, we expect the message structure and thus the
recognized types to change over time. Subsequently, we determine a dependency
between message types and activities dynamically through analysis of execution
traces. In a similar effort, Lakshmanan et al. [12] describe how an ant colony
optimization algorithm learns dependencies of document contents (e.g., the im-
pact of certain values within a message) to predict the flow and outcome of a
process. They also apply exponential aging to keep the decision probabilities up
to date. The underlying process model, however, remains unchanged.
Related work that explicitly applies autonomic computing principles [10] for
adaptive workflow and process support systems react to system internal events
such as workload fluctuations [8], goal changes [7], or service replacements [21]
when executing well-structured processes. These approaches, however, target
only system elements and offer no support for having the user dynamically adapt
the process. In contrast we aim to apply those autonomic principles primarily
for achieving unsupervised learning and user recommendation.
3 Approach
Successful activity recommendations in people-driven processes (i.e., the user
actually carries out the proposed activity) depend on correct classification of
incoming and outgoing messages and their associated activities. Supporting users
in dynamically evolving processes consists of two aspects: (i) run-time tracing of
process progress including probabilistic message prediction, and (ii) automatic
refinement of process model and message probabilities upon process termination.
The main phases of the run-time recommendation support is depicted in Fig-
ure 2. The core of the
Probabilistic Process Engine
consists of the process model
that describes the general structure and current interdependencies of steps and
messages (see Sec. 4.1). The Message State Model captures the likelihood for a
message to occur in that process on time, early, late, or repeatedly (see Sec. 4.2).
For each intercepted message and observed user action, the process engine
extracts the correlation of messages and activities to determine which activity
produced a particular message, and which message served as input to a given ac-
tivity (1). Next, the engine analyzes whether the observed messages and actions
