Service-oriented systems are deployed by companies to support business processes, especially in inter-organizational collaborative settings. Process mining provides techniques to visualize and understand the emergent behavior in such systems based on data, as compared to what employees believe it is like. In highly unstructured settings however, these techniques have to deal with incomplete data, which still is a demanding challenge. In order to address it, we propose an alternative “dual” interpretation of mining results and outline the theoretical basis upon which process mining techniques could be extended and modified to deliver such results.
With today’s pressure to rapidly adapt to highly dynamic business environments,
collaborations between different business partners have to be set up fast and in a
flexible manner. Ad-hoc formation of virtual teams is an increasingly observable
collaboration pattern. Technology can be considered one of the main facilitators of this
development [
For companies, it is important to be aware of their work practices in order to
manage and align their activities. Business processes, codified within models, constitute a
popular concept to do this [
Traditional modeling techniques such as Event-driven Process Chains (EPC) or the
Business Process Model and Notation (BPMN) are successful in representing
business processes of repetitive and well-standardized nature. Unstructured processes
though are seen as hardly amenable to traditional process modeling due to uncertainty
regarding their outcome as well as the steps and resources needed to produce it [
Research investigating process mining to the field of services includes technical
aspects such as logging in service-oriented architectures [
However, these works target well-structured processes. With respect to
collaborative service-oriented systems, little research has been done. Incompleteness,
describing a situation in which a mining algorithm is provided with a dataset not including all
necessary information, is a challenge process mining techniques must deal with. It is
of particular importance in collaborative, unstructured settings [
The question we want to investigate is if it is possible to adapt process mining techniques in a way such that they can be applied in settings in which event logs are expected to be far from complete. Naturally, we cannot expect to generate an exact description of the underlying real-world process. This raises the question what any process model generated on an incomplete dataset is supposed to represent. A possible answer can be obtained through the following line of reasoning.
The normal way of thinking about processes models is that they specify the allowed behavior of a system. As an example, consider a simple process in which activity A is performed first, and then either activity B or C is performed after that, which both terminates the process. This positive view defines the process as a set of possibly observable instances: { }. In such a setting, it makes sense to apply an algorithm to learn this allowed behavior. In a dual way though, one could equivalently think of a negative view that defines the process as the set of unobservable instances. In the example above, { } would be this set. While this set will be huge for highly structured processes, the opposite might be the case for unstructured ones. For this reason, algorithms searching for disallowed behavior in unstructured processes might use scarce data more efficiently.
But how can an algorithm infer disallowed behavior if only allowed behavior is observed in event logs? The answer is delivered by statistics. If direct negative evidence is unavailable one can work with indirect negative evidence instead, provided one is willing to assume a suitable statistical model. To illustrate this, consider the example of an unfair coin always showing heads when tossed. This could be inferred directly from the information that observing tails is the impossible event. Alternatively, one could observe that heads occurs surprisingly often, with each additional heads increasing the confidence that tails is a highly unlikely event. Statistical tests provide an established theoretical framework to incorporate such reasoning into process mining algorithms. Investigating how this can be accomplished is our approach to developing process mining techniques applicable to unstructured, collaborative serviceoriented systems.