Process mining techniques allow process analysts to generate process models from recorded event logs. Typically, process mining considers the event log as a whole and creates a single model re ecting its behavior. However, the process may be in uenced by several characteristics of the process instances, e.g., by the individual characteristics of a patient in the healthcare domain like age and sex. This leads to a wide range of process variations which can end up in complex and confusing models, blurring the behavior of speci c process variants. Multidimensional process mining (MPM) aims to overcome this limitation by the notion of data cubes, spreading the data over multiple cells, each representing a group of cases with similar characteristics. This allows for the creation of separated process models for a homogenous set of cases. In this paper, we introduce PMCube Explorer, a novel tool for MPM, that allows for the analysis of a process from various views. It enables the analyst to specify OLAP queries to extract multiple cells from the data warehouse. Each cell contains a subset of event data which are mined separately to discover independent process models. To deal with the potentially high amount of resulting models, our tool provides some distinctive features like the visualization of model di erences or the consolidation of multiple process models. We applied our tool in a case study to analyze the perioperative processes in a large German hospital.
Process mining is a set of techniques that allow analysts to generate process models from event logs which contain event data recorded during the execution of the process. However, the behavior of a process is often in uenced by several characteristics of the executed process instance. For instance, healthcare processes have to consider age, sex, and allergies of the patients. This leads to a wide range of process variations which can result in big and complex process models when analyzing them with process mining techniques. The notion of multidimensional process mining (MPM) aims to solve this problem by partitioning the underlying Copyright c 2015 for this paper by its authors. Copying permitted for private and academic purposes. 1 Multidimensional event log
Data selection (OLAP) Process mining
Consolidation
Visualization event log into subsets that consist of cases with homogenous features. These subsets (or sublogs) are mined separately to discover independent process models, each focusing on a limited feature combination of the cases. MPM adopts the concepts of OLAP and data cubes, that are commonly used in data warehouses (DWH), to the eld of process mining. The intention is to partition and lter the event log in a dynamic and exible way in order to provide customized views on the process.
In this paper, we demonstrate the PMCube Explorer, a novel tool for MPM. Section 2 brie y introduces the underlying concepts. In Section 3, we show the main features of our tool. A case study using the tool is discussed in Section 4. Finally, we brie y present the screencast of this demonstration in Section 5. 2
Figure 1 illustrates the PMCube approach, which is the underlying concept of
our tool. The multidimensional event log (MEL) is a DWH that stores the data
from the event log as a multidimensional data cube. In contrast to Event Cube
[
The result of an OLAP query is a set of independent sublogs. Process discovery techniques are applied to each sublog to discover a process model. Depending on the query, this may result in a high number of potentially complex process models which makes it hard for the analyst to interpret the results. To cope with this, PMCube provides an optional step of consolidation. It aims to reduce the number of process models by an automatic preselection of the most relevant process models. One consolidation approach is to cluster process models re ecting similar behavior and to select one representative process model per cluster. It is based on the heuristic that major di erences between process models are more relevant to the analyst than minor variations. After the consolidation, the results are visualized for interpretation. PMCube can arrange all process models side by side in a matrix to provide a general overview of the models. Alternatively, PMCube can also calculate the di erences between two models and highlight them in a merged model. 3
The PMCube Explorer is a prototypical implementation of the PMCube concept.
It is written in C# using the Microsoft .NET framework. The MEL is stored
in an external, relational database like Oracle or Microsoft SQL Server in an
advanced snow ake schema re ecting the two distinct levels for cases and events.
The analyst can query the MEL via a graphical user interface (GUI) to create a
customized view of the data. For this purpose, the GUI provides multiple options
to lter and aggregate the data cubes on both the case and the event level. For
each cell de ned by the OLAP query, a separated SQL query is created, which is
sent to the MEL. While executing the SQL queries, the multidimensional data
of the data cube is implicitly attened into a table where each line represents an
event of the sublog. Because the query results re ect the commonly used event
log structure, arbitrary process discovery algorithms can be used without any
adaptations to discover the process models. The resulting process models can be
visualized side by side in a matrix or as a single model. In contrast to other MPM
tools like Process Mining Cubes (PMC) [
The PMCube Explorer is highly extensible. All algorithms for process
discovery, conformance checking, di erence view calculation, and consolidation, as
well as the process model languages (data structures and view models) and the
database connectors are provided as plug-ins and loaded during run-time.
Currently, PMCube Explorer provides plug-ins for Inductive Miner { infrequent [
Figure 2 presents some screenshots of PMCube Explorer. They show the preview of the resulting cells while creating the OLAP query (1), the dialog for selecting the dimension that should be used for slicing (2), the matrix view (3), and the time perspective dialog showing the distribution of waiting times (4). 4
We conducted a case study where we used the PMCube Explorer to analyze healthcare processes of a university hospital as a center of maximum care in Germany. We started the evaluation study after the approval of the ethical committee of the Justus Liebig University (ethical review committee of the Faculty of Human Medicine at the Justus Liebig University Gie en, chairman Prof. Dr. Tillmanns, vote number 261/14) with an anonymized data set. This data set covers a random sample of 16,280 surgical interventions of four medical departments in 2012 and 2013 with a total of 388,395 events. We focused on the perioperative process, which comprises all activities in the periphery of surgical interventions, especially activities related to anesthesia. The event data was extracted from several clinical information systems and anonymized by the hospital IT, before we integrated it into the multidimensional structure of the MEL.
We applied multiple OLAP queries in an explorative way to analyze the processes from various points of view. We discussed the discovered models with a medical expert, who is familiar with the perioperative processes of that hospital. The case study showed that MPM provides a dynamic and exible way to analyze processes from di erent views. Queries can be easily adjusted, which allows for the explorative analysis of the processes. However, the case study revealed that MPM can become quite complex and confusing, especially when comparing many process models. Although the consolidation and the di erent visualization techniques showed to be a helpful tool during analysis, they need to be improved and complemented by more sophisticated techniques to deal with the high complexity of results. This should be tackled by future work. 5
A screencast that demonstrates the usage of the PMCube Explorer is available on the web (http://youtu.be/CTXyIZp2BJw). It gives a walk-through of an example process mining analysis, conducted on the data of the case study described in Section 4. Starting with the creation of an OLAP query, it presents the main features of the tool, like the clustering-based consolidation of process models, the matrix visualization, and the automatic visualization of di erences between process models. Furthermore, it shows how to switch the view of the data cube. Acknowledgments. The authors would like to thank all contributors to the PMCube Explorer and especially Rainer Rohrig, Lena Nieho , Raphael W. Majeed, and Christian Katzer for their support during the case study.