Today’s business processes are realized by a complex sequence of tasks that are performed throughout an organization, often involving people from different departments and multiple IT systems. For example, an insurance company has a process to handle insurance claims for their clients, and a hospital has processes to diagnose and treat patients. Because there are many activities performed by different people throughout the organization, there is a lack of transparency about how exactly these processes are executed. However, understanding the process reality (the "as is" process) is the first necessary step to save cost, increase quality, or ensure compliance. The field of process mining aims to assist in creating process transparency by automatically analyzing processes based on existing IT data. Most processes are supported by IT systems nowadays. For example, Enterprise Resource Planning (ERP) systems such as SAP log all transaction information, and Customer Relationship Management (CRM) systems are used to keep track of all interactions with customers. Process mining techniques use these low-level log data (so-called event logs) to automatically generate process maps that visualize the process reality from different perspectives. For example, it is possible to automatically create process models that describe the causal dependencies between activities in the process. So far, process mining research has mostly focused on the discovery aspect (i.e., the extraction of models from event logs). This dissertation broadens the field of process mining to include the aspect of conformance and extension. Conformance aims at the detection of deviations from documented procedures by comparing the real process (as recorded in the event log) with an existing model that describes the assumed or intended process. Conformance is relevant for two reasons: 1. Most organizations document their processes in some form. For example, process models are created manually to understand and improve the process, comply with regulations, or for certification purposes. In the presence of existing models, it is often more important to point out the deviations from these existing models than to discover completely new models. Discrepancies emerge because business processes change, or because the models did not accurately reflect the real process in the first place (due to the manual and subjective creation of these models). If the existing models do not correspond to the actual processes, then they have little value. 2. Automatically discovered process models typically do not completely "fit" the event logs from which they were created. These discrepancies are due to noise and/or limitations of the used discovery techniques. Furthermore, in the context of complex and diverse process environments the discovered models often need to be simplified to obtain useful insights. Therefore, it is crucial to be able to check how much a discovered process model actually represents the real process. Conformance techniques can be used to quantify the representativeness of a mined model before drawing further conclusions. They thus constitute an important quality measurement to effectively use process discovery techniques in a practical setting. Once one is confident in the quality of an existing or discovered model, extension aims at the enrichment of these models by the integration of additional characteristics such as time, cost, or resource utilization. By extracting aditional information from an event log and projecting it onto an existing model, bottlenecks can be highlighted and correlations with other process perspectives can be identified. Such an integrated view on the process is needed to understand root causes for potential problems and actually make process improvements. Furthermore, extension techniques can be used to create integrated simulation models from event logs that resemble the real process more closely than manually created simulation models. In Part II of this thesis, we provide a comprehensive framework for the conformance checking of process models. First, we identify the evaluation dimensions fitness, decision/generalization, and structure as the relevant conformance dimensions.We develop several Petri-net based approaches to measure conformance in these dimensions and describe five case studies in which we successfully applied these conformance checking techniques to real and artificial examples. Furthermore, we provide a detailed literature review of related conformance measurement approaches (Chapter 4). Then, we study existing model evaluation approaches from the field of data mining. We develop three data mining-inspired evaluation approaches for discovered process models, one based on Cross Validation (CV), one based on the Minimal Description Length (MDL) principle, and one using methods based on Hidden Markov Models (HMMs). We conclude that process model evaluation faces similar yet different challenges compared to traditional data mining. Additional challenges emerge from the sequential nature of the data and the higher-level process models, which include concurrent dynamic behavior (Chapter 5). Finally, we point out current shortcomings and identify general challenges for conformance checking techniques. These challenges relate to the applicability of the conformance metric, the metric quality, and the bridging of different process modeling languages. We develop a flexible, language-independent conformance checking approach that provides a starting point to effectively address these challenges (Chapter 6). In Part III, we develop a concrete extension approach, provide a general model for process extensions, and apply our approach for the creation of simulation models. First, we develop a Petri-net based decision mining approach that aims at the discovery of decision rules at process choice points based on data attributes in the event log. While we leverage classification techniques from the data mining domain to actually infer the rules, we identify the challenges that relate to the initial formulation of the learning problem from a process perspective. We develop a simple approach to partially overcome these challenges, and we apply it in a case study (Chapter 7). Then, we develop a general model for process extensions to create integrated models including process, data, time, and resource perspective.We develop a concrete representation based on Coloured Petri-nets (CPNs) to implement and deploy this model for simulation purposes (Chapter 8). Finally, we evaluate the quality of automatically discovered simulation models in two case studies and extend our approach to allow for operational decision making by incorporating the current process state as a non-empty starting point in the simulation (Chapter 9). Chapter 10 concludes this thesis with a detailed summary of the contributions and a list of limitations and future challenges. The work presented in this dissertation is supported and accompanied by concrete implementations, which have been integrated in the ProM and ProMimport frameworks. Appendix A provides a comprehensive overview about the functionality of the developed software. The results presented in this dissertation have been presented in more than twenty peer-reviewed scientific publications, including several high-quality journals.
|Qualification||Doctor of Philosophy|
|Award date||3 Nov 2010|
|Place of Publication||Eindhoven|
|Publication status||Published - 2010|