We present our eforts in creating and analyzing a corpus of BPMN process models by mining software repositories. Systematically searching for BPMN process artifacts in 6,163,217 repositories or 10% of all repositories hosted on GitHub.com, at the time of conducting our research, resulted in a diverse corpus of 8,904 BPMN 2.0 process models. Within the last years, an increasing number of software projects have shifted towards using platforms such as GitHub.com for their software development. Using these platforms as a source of data for empirical research allows for addressing a wide range of questions on the practice of software development and receives more and more attention, as indicated by the popularity of the flagship conference on the topic: International Conference on Mining Software Repositories (MSR)1. Research in the domain of business process modeling can as well benefit from such a data-driven approach. Due to characteristics of the domain, i.e., “process equals product”, there is a lack of larger and commonly available datasets with real-world process models, which hinders empirical research in this area [2,11,13]. Mining software repositories, i.e., systematically retrieving, processing and analyzing process models from software repositories hosted on platforms such as GitHub.com, can help to overcome this lack and provides a complimentary approach to empirical research besides existing methods like case studies, experiments, and surveys. For example, research questions on how a language such as the Business Process Model and Notation (BPMN) [1] is used in practice can be addressed, in order to diferentiate between the frequently and the rarely used parts of the language, thus advancing language and tool development. Analyzing modeling styles furthermore allows for investigating best practices and guidelines to help process designers. Eventually, best practices and tools as proposed by academic research or industry can be evaluated more realistically [12]. In this paper, we present our approach for mining software repositories on GitHub.com to create and analyze a corpus of BPMN process models. Due to the sheer number of repositories on GitHub.com and time constraints, we limited our approach to a randomly selected subset of 6,163,217 repositories or 10% of 1 http://www.msrconf.org
GHTorrent
GitHub
GitHub
API v3 1 Repository
Selection 2 Data
Extraction 3 Filtering /
Cleansing 4 Analysis all software repositories on GitHub.com at the time of conducting the research. As a result, we were able to identify and analyze 8,904 distinct process models which are defined using BPMN 2.0’s XML-based serialization format. 2
The Lindholmen dataset has been an inspiration for this paper [
There have also been community eforts to create model collections [
Mining software repositories is a data mining task, consisting of steps of defining a research objective, selecting and extracting appropriate data, preprocessing and data cleansing, data analysis, and finally interpreting the analysis results.
In the first step of our implemented data mining pipeline, compare with Fig. 1, we got a list of all software repositories on GitHub.com by querying a local instance of the GHTorrent2 database. We then randomly selected a subset
Thomas S. Heinze et al.
of 6,163,217 non-forked repositories. All 6,163,217 repositories were examined for
potential BPMN process model artifacts using the GitHub API 3 in the second
step. To this end, the default branch and its file structure were queried for
each repository. Potential BPMN process model artifacts were then identiefid
by searching for the term "bpmn" in their file name and file extension. Among
the analyzed repositories, we found 1,251 repositories, with at least one potential
BPMN process model artifact and overall 21,306 artifacts. We downloaded the
identified repositories and artifacts. In the third step, since the artifacts included
a wide range of file formats, we filtered for BPMN 2.0’s XML-based serialization
format, which lowered the number of artifacts to 16,907. Additionally removing
duplicates4, yielded the corpus of 8,904 distinct BPMN 2.0 models. All the
BPMN artifacts were finally subject to a preliminary analysis in the fourth step.
Information on the corpus and analysis outcomes are available online [
In our preliminary analysis, we were mainly interested in the diversity of the
found BPMN process model artifacts. We here sketch some of the results. Looking
at the artifacts’ age, more than each third was modified in the last year at the
time of conducting our research. We though also found artifacts older than 8
years. Using the locations of repository contributors allowed us to reason on the
artifacts’ geographical origin, where China, USA, and Germany played prominent
roles. The corpus spans a range of diferent model sizes. While half of the process
models are smaller than 20 nodes, we also identified 57 models with more than
1,000 nodes. We were also able to confirm the finding reported in [
Since the design of BPMN process models is known to be error-prone, we
were also interested in the number of errors found in the models and the need
for analysis tools to help process designers in avoiding those. Various analysis
tools have been developed in recent years, ranging from simple linters [
In this paper, we introduced our approach of systematically extracting a corpus of
BPMN business process models from software repositories hosted on GitHub.com.
Mining a fraction of 10% of all software repositories, at the time of conducting our
research, resulted in 8,904 distinct serialized BPMN 2.0 process models. We believe
that our corpus of BPMN models provides a starting point for understanding
more about the practice of BPMN. Note though the general limitations of the
idea of repository mining [