With the prevalence of business process management in the private and public sector, large process collections are created and shift into focus. To be able to harvest the underlying information, process collections need to be made easily accessible providing intuitive navigation and search. Process collections are often structured into folders that are labeled along functional, organizational or goal-based lines. As those structures are tediously created manually, they often offer only a single view on the underlying data. However, users use such process collections with different intentions. This paper presents a generic approach for automatically creating process architectures from unstructured process collection to offer browsing and user centered navigation structures, as well as reduce time of creation. The approach uses the characteristics of clustering algorithms to group processes and label them accordingly. Improvements for further development in the near future will be investigated and outlined as well.
In recent years BPM has gained momentum in the private and public sector. As
a result, process collections of different size, quality, and purpose have emerged.
Especially for large collections, the need for an inherent and intuitive structure
and navigation is of importance for the retrieval of process models. The knowledge
stored in process collections is often not treasured to the best possible extent.
The lack of consistent ordering poses strong challenges for the retrieval of process
models. According to Baeza-Yates and Ribeiro-Neto [
Offering structured overview of business processes in a process collection is
one of the aims of process architectures according to Dijkman et al. [
In well-organized projects, the process elicitation process follows guidelines for structuring process models according to a process architecture. However, in many cases, process architectures have not been developed before the modeling phase. As a result, many process collections are semi- or unstructured.
(Re-) structuring already existing process collections becomes a strenuous
manual task, starting with the design of a process architecture. Later, processes
need to be classified manually into specific categories. For collections of
hundreds or several thousand process models, e.g., SAP Reference Model, Dutch
administration, or China CNR Corporation Limited [
This paper will not focus on defining process architectures in the beginning of business process modeling projects in regard to modeling responsibilities, guidelines, undocumented processes or other issues of process architecture design. It rather presents a generic system design and algorithm architecture along with a concrete example that creates a process architecture based on syntactic similarity of process names. This approach shall provide better navigation through process knowledge, as well as improve efficiency and adequacy over the manual creation of process collection structures. It may even bear the possibility to create process architectures with different focus according to the users’ interest.
The paper is structured as follows, Sect. 2 will introduce current research on the structuring of process collections, process architectures and hierarchical clustering, Sect. 4 presents a conceptual system architecture, Sect. 5 sketches an algorithm for creating process architectures, followed by Sect. 6 which will elaborate on future work and improvements of the presented ideas. 2
Different approaches to structuring process collections or creating process
architectures have been developed and proposed. Most of them are based on manual
classification techniques. Weske [
Dijkman et al. [
Scheer et al. [
In Smirnov et al. [
Coming from a different domain but facing similar problems in large
multimedia collections Lew et al. [
Qiao et al. [
Jung et al. [
Most of the approaches use similarity measure and clustering algorithms to
find similar process models or similar process elements focusing on structural [
The reduction of complexity and the visualization of the most important
characteristics of process models is the aim of many approaches. Abstraction,
as well as providing browsing and summarization of process information can
be achieved by creating process architectures. According to Baeza-Yates and
Ribeiro-Neto [
Most of the process architecture styles presented in [
Hierarchical clustering algorithms can be distinguished into agglomerative
clustering (bottom-up) and divisive clustering (top-down). Agglomerative
clustering algorithms are single-link, average-link, complete-link and centroid-based
which differ on their similarity measures. Divisive clustering (top-down) starts
with one cluster that iteratively becomes split into smaller clusters. In contrary
to agglomerative clustering, divisive clustering is more complex and uses flat
clustering algorithms, like k-means clustering in its subroutine [
In van Rijsbergen [
Despite that, the use of agglomerative single link clustering (hierarchical clustering) is promising for automatically creating process architectures considering its simplicity and robustness. It fulfills the adequacy requirements making it easily applicable to heterogeneous as well as different process collections. The drawback of exponential computational complexity can be disregarded as creating process architectures is not an everyday task. 4
to the selected cluster algorithm. This shall allow exploring the ability of relating
clustering algorithms to particular user views. The results of the clustering process
are clusters arranged in a hierarchical tree structure. The resulting hierarchy tree
diagram is depicted as dendogram [
After the clustering process, the label analyzer module analyzes each member of a cluster and chooses a label for the cluster considering common characteristics of process models e.g. metadata, labels, or input and output. Algorithms for selecting adequate labels for clusters using semantic approaches from the field of natural language processing will be part of future work. 5
The general approach for creating process architectures from process collections, shown in Fig. 2, consists of three steps; preprocessing, clustering, and labeling of clusters, which will be presented along with a concrete example Fig. 3. The approach takes a model collection as input and creates a process architecture with labeled clusters as output. In the preprocessing step, process models and their metadata are cleansed. Metadata values are normalized and missing data is dealt with. String values are converted into numerical values. There are different approaches that can be applied to deal with missing data and string values of metadata. In the example algorithm the preprocessing step consists of extracting the process names from the process models and converting the strings into numerical values by calculating the string edit distance. The output of the step is multidimensional vectors representing the process models. In future research this step could be improved by using natural language processing techniques for dealing with semantics of process names, or labels of control flow elements. For example synonyms could be detected and given the same value. Analyzing the semantic similarity of labels of control flow elements, structural, or behavioral aspects of process models bears many possibilities for exploration. Particular process model elements shall be aligned to styles of process architectures and form the input for generating those automatically; e.g. activities could be used to generate action-based process architectures. The cluster function takes the Function PreprocessProcessModels; Input ProcessCollection PC, List SelectionOfMetadata; Output ListPreprocessedModels PM ; Function Cluster; Input PreprocessedModels PM, ClusterAlgorithm CA, SimFunction SF ; Output Dendogram D, SetClusters C ; Function ClusterLabeling; Input Dendogram D, SetClusters C, AnalysisAlgorithm AA ; Output LabeledDendogram LD, SetOfLabeledClusters LC ; Example: Function PreprocessProcessModels; Input SAPReferenceModel, Processname; Output multVectors Example: Function HierarchicalBottomUpCluster; Input multVectors, HierachicalSLBottomUp, EuclideanDistance ; Output (see Fig. 4(a)) Dendogram, Clusters Example: Function ClusterLabeling; Input Dendogram, Clusters, SimNamePartsAndCountEvents ; Output (see Fig. 4(b)) LabeledDendogram, labeledClusters
Function General and Example Function Interface Descriptions preprocessed process models, e.g., a multidimensional vector as input as well as the clustering algorithm of interest and a similarity function to calculate the similarity between the different processes in the process collection. The output is a dendogram and a set of clusters.
The input of the example algorithm is a list of multi-dimensional vectors representing the process models, the bottom-up single-link clustering algorithm and the Euclidean distance function. A hierarchical clustering algorithm is chosen due to the hierarchical nature of most process architectures and the good characteristics of hierarchical clustering algorithms. The Euclidean distance function is used to calculate the similarity matrix for the process models. In future, this step can be realized with more sophisticated similarity functions. Considering each process model as one cluster in the beginning, the hierarchical cluster algorithm will join the two clusters with the minimal distance between any two items pi and pj with pi in cluster i and pj in cluster j until only one cluster exists containing all other clusters and the inherent process models. After each iteration, the distance matrix must be updated with the similarity value of the newly created cluster. The clustering process results in a dendogram, a hierarchical structure tree that links all the clusters generated.
In the future also flat clustering algorithm can be used with presented framework. The next step defines the labeling process of the clusters. The general algorithm takes the dendogram, the set of clusters and an analysis algorithm as input. Here different strategies that need further elaboration can be applied. E.g. only the input and output labels of each member in the cluster could be extracted and counted.
In the example, the analysis algorithm examines the names of each member of cluster and figures out a word that describes the processes in the cluster. It also counts the number of activities, start events, and end events of each process. The label is put together from a word that is common for each process model in the cluster as well as the range of start events, end events and activities as displayed in Fig. 4(b). In this way context information on the process models in each cluster is provided while browsing. The technique described in the example is rather a simple technique for identifying labels which also leaves room for improvement in the future.
An example output of the clustering algorithm with five process models from the sap reference model collection is depicted in Fig. 4(a) and Fig. 4(b). The (a) (b) implementation of the algorithm has not been fully realized and first results with large process collections can only be presented in the near future.
Empirical surveys can be used to validate both the automatically created
process architectures in regard to their improvement of browsing capabilities and
the labels chosen for the different clusters. The architectures can be replicated
in process collections and tested with users to investigate the quality of their
browsing capabilities. In a similar way, the quality of labels representing the
process clusters can be assessed by users in an empirical survey. A current
research project, the national process library (npl), offers a suitable use case for
this purpose [
This paper presented a conceptual framework for automatically generating process architectures from process collections. A general and exemplary algorithm as well as the input and output of the different steps were presented to depict the process of generating process architectures for process collections. Suggestions for improvements of the quality of clustering and labeling of clusters were mentioned as future research agenda. Using more sophisticated similarity measures for clustering as well as natural language processing techniques for analyzing semantic information from control flow labels may bear the most potential here. Also, the preprocessing step can be varied in respect to semantics which likely improves results, e.g. only nouns will be extracted. In general the presented approach is very flexible and lays the foundation for future exploration of process collections and the interdependencies of its process models.