Modelling Business Process Variants using Graph Transformation Rules
Christine Natschl
¨
ager
1
, Verena Geist
1
, Christa Illibauer
1
and Robert Hutter
2
1
Software Competence Center Hagenberg, Hagenberg, Austria
2
Prologics IT GmbH, Linz, Austria
Keywords:
Business Process Management, Variability Modelling, Variant Management, Graph Transformation.
Abstract:
Business process variability is an active research area in the field of business process management and deals
with variations and commonalities among processes of a given process family. Many theoretical approaches
have been suggested in the last years; however, practical implementations are rare and limited in their func-
tionality. In this paper, we propose a new approach for business process variability based on well-known graph
transformation techniques and with focus on practical aspects like definition of variation points, linking and
propagation of changes, as well as visual highlighting of differences in process variants. The suggested con-
cepts are discussed within a case study comprising two graph transformation systems for generating process
variants; one supports variability by restriction, the other supports variability by restriction and by extension.
Both graph transformation systems are proven to be globally deterministic, but differ regarding their complex-
ity. The overall approach is being implemented in the BPM suite of our partner company.
1 INTRODUCTION
Process variability is related to flexible business pro-
cess management (BPM), one of the most active re-
search areas (Reichert and Weber, 2012). In de-
tail, process variability is concerned with design- and
customization-time decisions and deals with differ-
ences and commonalities among processes of a given
process family (Rosa et al., 2013). Typical fields of
application are, e.g., production processes, invoice
processes, and delivery processes.
There is a large number of theoretical approaches
to process variability, focusing on variability by re-
striction and/or by extension. However, many experi-
enced researchers assess the tool support for handling
process variants as limited and not adequate (Reichert
and Weber, 2012; Rosa et al., 2013). In addition,
exponential growth makes modelling and maintain-
ing process variants a complex endeavour in practice.
Thus, there is need for establishing a sound method
The research reported in this paper has been partly
supported by the Austrian Ministry for Transport, Innova-
tion and Technology, the Federal Ministry of Science, Re-
search and Economy, and the Province of Upper Austria
in the frame of the COMET center SCCH. This publica-
tion has been written within the project AdaBPM (number
842437), which is funded by the Austrian Research Promo-
tion Agency (FFG).
for explicitly supporting variability in business pro-
cesses in a way that fosters industrial applicability.
In this paper, we propose a new concept for vari-
ant management based on graph transformation tech-
niques. Important aspects of this concept are the indi-
vidual definition of adaptable and blocked elements as
well as linking and propagation of changes. A further
advantage is the application of graph transformation
rules to deal with the exponentially increasing num-
ber of variants. Instead of manually defining all vari-
ants, a few rules specify concrete variations and the
corresponding variants are automatically generated.
The paper is structured as follows. In Section 2,
we provide an overview of related work on graph
transformation, variability modelling, and tool sup-
port. We present the general approach for modelling
business process variants based on key requirements
in Section 3 and discuss the proposed concepts within
a case study using graph transformation techniques in
Section 4. In Section 5, we summarise our findings.
2 STATE-OF-THE-ART
This section provides an overview of the state-of-the-
art concerning graph transformation, variability mod-
elling, and the current support of business process
variability in BPM suites.
Natschläger, C., Geist, V., Illibauer, C. and Hutter, R.
Modelling Business Process Variants using Graph Transformation Rules.
DOI: 10.5220/0005665800650074
In Proceedings of the 4th International Conference on Model-Driven Engineering and Software Development (MODELSWARD 2016), pages 65-74
ISBN: 978-989-758-168-7
Copyright
c
2016 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
65
2.1 Graph Transformation
Research on graph transformations started around the
1970s; the main idea is the rule-based modification of
a graph, where each application of a rule leads to a
graph transformation step. So the transformation of a
graph L (left-hand side) to a graph R (right-hand side)
is based on a rule r (also called production). Apply-
ing the rule r = (L, R) means finding a match of L in
the source graph and replacing L by R, leading to the
target graph of the graph transformation. The appli-
cation of rules is restricted by application conditions
(AC), which can either be positive or negative. A neg-
ative application condition (NAC) is satisfied if it is
not part of a match of L in graph G. A positive ap-
plication condition (PAC) is the counterpart to a NAC
and is satisfied if it is part of the match of L in G
(Ehrig et al., 2006, p. 5f, 67ff).
Different approaches for graph transformation
have been proposed like node label replacement,
hyperedge replacement, or the algebraic approach
(Ehrig et al., 2006, p. 10). The algebraic approach,
which was initiated by Ehrig, Pfender, and Schneider
in 1973, is applied for the suggested variability con-
cept. Algebraic graph transformation is supported by
a tool for attributed graph grammar systems (called
AGG). AGG provides a graphical editor and can be
used for specifying graph grammars with a start graph
or for typed attributed graph transformations. In ad-
dition, AGG offers analysis techniques as for consis-
tency checking, critical pair analysis, and termination
evaluation (Ehrig et al., 2006, chapter 15). In previous
work, AGG was applied to implement an algebraic
graph transformation from standard BPMN to Deon-
tic BPMN diagrams and to prove that this is a trusted
model transformation (Natschl
¨
ager et al., 2015). In
Section 4, AGG is used to implement and evaluate
the suggested variability concept.
The algebraic approach to graph transformation
has been generalised to further graph types and high-
level structures such as typed graphs, labelled graphs,
hypergraphs, attributed graphs, and algebraic speci-
fications. This extension to high-level structures is
called high-level replacement (HLR) (Ehrig et al.,
2006, p. 14ff). The graph transformation for the pro-
posed variability concept is defined as typed attributed
graph transformation system with inheritance that in-
cludes labels.
2.2 Variability Modelling
A survey on business process variability modelling
is provided in (Rosa et al., 2013). The authors dis-
tinguish three phases in the lifecycle of customiz-
able process models: design-time, customization-
time, and runtime. Variability is concerned with
design- and customization-time decisions and subdi-
vided into variability by restriction (i.e., customiz-
able model is the union or least common multiple
of all process variants) and variability by extension
(i.e., customizable model is the intersection or great-
est common denominator of all process variants). The
authors further argue that there is no explicit support
for representing and maintaining multiple variants of
a defined business process in conventional business
process modelling languages. Available approaches
typically extend a modelling language with additional
constructs to capture customizable process models.
Variability by restriction is applied by so-called
configurable approaches. Configurable approaches
require a comprehensive reference model that com-
prises all possible process flows and which is config-
ured at design-time to the required variants (see (Re-
ichert and Weber, 2012) for details). In (Aalst,
2013), twenty BPM use cases are suggested, three of
them relate to configurable process models: design
configurable model, merge models into configurable
model, and configure configurable model. Business
process modelling languages supporting the config-
urable approach are, for example, Configurable YAWL
(C-YAWL) and Configurable EPCs (C-EPCs). In
C-YAWL, hiding and blocking operations are applied
to the input and output ports of activities, so activ-
ities can either be hidden (skipped) or the trigger-
ing of the activity (and the subsequent path) is pre-
vented (blocked) (Gottschalk, 2009). C-EPCs, on the
other hand, provide configurable functions (may be
included, excluded, or conditionally skipped), con-
figurable connectors (can be mapped to more restric-
tive connectors), and configurable requirements and
guidelines (If-Then-statements to define dependencies
between configurable nodes) (Recker et al., 2006).
Configurable approaches are useful to adapt, e.g., a
domain-specific reference model to a concrete organ-
isation. However, the underlying modelling language
must be extended with configurable elements and the
reference model is extensive since it must comprise
all process flows. Furthermore, local process variants
are restricted in their adaptability.
Adaptable Approaches, in contrast, only require
a base process model, which comprises the standard
process flow and can be adapted through structural
model adaptations based on change patterns (vari-
ability by restriction and by extension). In (We-
ber et al., 2008), eighteen change patterns are sug-
gested and divided into adaptation patterns (e.g.,
add/delete/move/replace fragment) and patterns for
changes in predefined regions to deal with uncer-
MODELSWARD 2016 - 4th International Conference on Model-Driven Engineering and Software Development
66
tainty. In addition, (Doehring et al., 2011) propose
a pattern catalogue with patterns for basic adapta-
tions, time adaptation and exception handling. The
adaptable approach is implemented, e.g., by Provop.
Provop is an approach for modelling and managing
process variants, covering the whole process lifecy-
cle. In the modelling phase, a base process is de-
fined, comprising adjustment points to restrict the re-
gions to which adaptations may be applied. In ad-
dition, a set of reusable change operations is spec-
ified, which can be grouped into so-called options.
Provop supports the following change patterns: in-
sert/delete/move fragment and modify attribute. In
the configuration phase, the user selects a sequence of
options and applies them to the process to configure
the desired process variant. The configuration may be
context-aware, meaning that the selection and appli-
cation of options is based on the context. The con-
text model then comprises a set of context variables
as well as restricting context constraints. Each con-
text variable specifies one particular dimension, be-
ing visualised in a context cube in which every cell
represents a variant. In the execution phase, a config-
ured variant is transformed into an executable work-
flow model and deployed to the runtime environment.
Runtime flexibility is provided by the possibility to
dynamically switch execution between different pro-
cess variants. Finally, in the optimization phase, the
base process may be changed according to previous
adaptations, so that the process family evolves over
time (Hallerbach et al., 2009; Hallerbach et al., 2010).
The main similarity between the Provop approach
and our variability concept is that both approaches ap-
ply variability by restriction and by extension. Dif-
ferences concern the implementation of adaptations
(concept of change operations and options vs. graph
transformation techniques with formal foundation)
and the restriction of adaptations (adjustment points
vs. individual blocked/adaptable elements).
Summing up, none of the proposed configurable
or adaptable approaches fits for our purpose. They
either introduce additional constructs to a modelling
language in order to support process variability, re-
quire extensive reference models, restrict adaptations,
or propose concepts without formal foundation.
2.3 Variability Modelling in BPM Suites
We studied the practical application by comparing in
detail six different BPM suites regarding their sup-
port for variability modelling and variant manage-
ment. The first BPM suite is called Aeneis and, ac-
cording to its documentation, claims to support vari-
ant management (Aeneis, 2014). Variant management
is, however, implemented by a simple copy & paste
mechanism. The variant (i.e., copy of a process) can
be modified without restrictions, since elements are
by default not linked to corresponding elements of the
parent process. Links can be inserted manually, but
then all linked elements are identical and neither side
can be changed without propagation of the change to
the other element. Processes (and variants) can be
compared by a basic comparison tool that lists all el-
ements of each process without highlighting any dif-
ferences. So, in our opinion, Aeneis does not provide
actual variability modelling, but it offers some func-
tionalities required by a rudimentary variant manage-
ment system.
The BPM suite Signavio (Signavio, 2015) does
not mention the support of variant management in
its documentation but provides appropriate function-
ality. First of all, Signavio supports rudimentary vari-
ant management by copying processes and permitting
to adapt the copy (no linking). A graphical compar-
ison tool identifies all changes between two process
models except swap operations. In addition, Signavio
supports stakeholder-specific views that resemble a
variant management with a reference model. After
having specified the overall process model (i.e., refer-
ence model), views are defined by selecting and omit-
ting elements (variability by restriction). Only the set-
tings of a view are saved and the view is generated
whenever required. Graphical model comparison is
provided between the reference model and the view
highlighting all differences in the reference model.
Scheer Business Process as a Service (BPaaS)
(Scheer, 2015) offers a simple variant management
limited to variability by restriction with the Scheer
Process Tailor. A reference process is defined and
variants are created by copying the reference process
and removing unnecessary activities. Thus, adapta-
tion of local variants is restricted to delete operations.
Three further BPM suites only support copy &
paste of a process model, which is insufficient for
variant management. Bizagi with its products Mod-
eler, Studio, and Engine is a comprehensive BPM
suite, also including a simulation component but no
variant management (Bizagi, 2014). Axon IVY, based
on the eclipse platform, provides six modules for
modelling, publication, analysis, design, execution,
and monitoring but again variant management is miss-
ing (Axon IVY, 2015). Lastly, BPM suite FireStart
of our industrial partner (Prologics, 2014) currently
lacks support for variant management but is being ex-
tended with the suggested variability concept.
In addition, to the best of our knowledge, also
other well-known tools such as the MS Workflow
Manager, IBM Business Process Manager, jBPM,
Modelling Business Process Variants using Graph Transformation Rules
67
and the BPM suites by Bonitasoft, Camunda, Inubit,
TIBCO, K2, and Appian do not fulfill our require-
ments for variability modelling.
Summing up, although variability modelling and
variant management have been discussed in detail
from a theoretical perspective, practical applications
are very limited. Only two of the analysed BPM
suites (Signavio and Scheer BPaaS) support variabil-
ity by restriction and thus part of variant management.
Current implementations especially lack functionality
regarding connections between parent and child pro-
cesses (i.e., links), propagation of changes, and vi-
sual highlighting of differences. These issues are ad-
dressed by the following variability concept.
3 A CONCEPT FOR
VARIABILITY MODELLING
In this section, we present our concept for variabil-
ity modelling, which is called Adaptive Variant Mod-
elling (AdaVM) and part of the overall AdaBPM
framework for advanced adaptivity and exception
handling in formal business process models. The con-
cept of AdaVM is based on the following key require-
ments for business process variability of our part-
ner company Prologics, producer of the BPM suite
FireStart. The requirements stem from a comprehen-
sive BPM vendor survey (2014):
1. definition of a main process and corresponding
variants,
2. local adaptation/extension of variants,
3. visual highlighting of differences between main
process and variants, and
4. propagation of changes in main process to vari-
ants with notification of responsible person.
So, at first, the main process model (also called
base or default process model) is defined like any
other process model, with two exceptions: The first
exception is that AdaVM restricts variation points by
providing Boolean attributes to globally block insert
operations (to support variability by restriction) and
to locally block modify and delete operations of in-
dividual elements or attributes. Blocked elements or
attributes indicate parts common to all variants and
can neither be changed nor deleted. The second ex-
ception is that the main process model may not be
a semantically valid process variant. Five different
policies have been suggested in (Reichert and Weber,
2012, p. 104f) for specifying the main process model:
(1) reference process, (2) most frequently used pro-
cess, (3) minimum adaptation efforts, (4) superset of
all process variants, and (5) intersection of all process
variants. Only the first and second policies represent
a particular variant and are directly executable. The
main process model of the fourth policy can only be
adapted by delete operations (configurable approach).
AdaVM does not prescribe a policy but it supports and
recommends to use both techniques, variability by re-
striction and by extension (adaptable approach) (see
discussion in Section 4).
In the following, we describe the general approach
of AdaVM for modelling business process variants.
A derived variant is at the beginning a duplication
of the main process model. Every element in the
variant is linked to the corresponding element in the
main process model, e.g., by the same ID or an ad-
ditional attribute specifying the ID of the parent el-
ement. Blocked elements are marked with a block
symbol and disabled (see Figure 1(a)), so that they
can neither be changed nor deleted. All unblocked el-
ements of the variant represent variation points and
can be adapted/extended by applying the following
change operations: insert, delete, and modify. Change
operation move suggested by Provop in (Hallerbach
et al., 2010) is substituted by deletion and insertion of
edges. Three symbols are introduced and correspond
to the three change operations. An inserted element
(node or edge) is marked with a plus symbol (see Fig-
ure 1(b)) and a modified element is marked with a
pencil symbol (see Figure 1(c)). A deleted element,
however, is usually removed from a process model,
thereby losing any graphical information about the
change. Thus, we decided to disable and grey-out
deleted elements and mark them with a cancel symbol
(see Figure 1(d)). Advantages of this approach are the
visibility of the change and the possibility to restore
a deleted element with its link to the corresponding
parent element. So the link to the corresponding ele-
ment is maintained even if one or more attributes are
changed or the whole element is deleted, only new el-
ements have no connection. An important aspect of
the suggested concept is that also edges are linked to
their parent edges. This provides the possibility to
identify moved (or swapped) elements, a change that
remains unrecognised by version comparison of many
well-known BPM suites (e.g., Signavio or Aeneis).
(a) Blocked (b) Added
(c) Modified (d) Deleted
Figure 1: Marked Elements.
MODELSWARD 2016 - 4th International Conference on Model-Driven Engineering and Software Development
68
The number of variants derived from the main pro-
cess model is not restricted. In some cases, however,
a desired process variant has more similarities (and
dependencies) with an existing variant than with the
main process model. AdaVM addresses this issue by
supporting single inheritance between variants, i.e.,
a variant is derived from another variant. The re-
sult is a tree-structured hierarchy of variants with the
main process model as the root (see Figure 2). Inher-
ited elements of a derived variant are always linked
to the corresponding elements of its parent process
model (e.g., elements of variant V2.1 are linked to
elements of variant V2 which are in turn linked to
elements of main process model MPM). In addition,
every process model passes on all inherited blocking
constraints without modification to its derived process
models and may add further restrictions. For example,
variant V2 inherits blocking constraints from the main
process model MPM and may define additional block-
ing constraints. All blocking constraints are passed
on to the derived variants, i.e. V2.1, V2.2 (and subse-
quently to V2.2.1).
Figure 2: Hierarchy of Main Process Model and Variants.
The last requirement addressed by AdaVM is the
propagation of changes in the main (or parent) pro-
cess to derived variants with notification of the re-
sponsible person. Since variability is concerned with
design- and customization-time decisions, no runtime
adaptations are expected. The responsible person of
a derived variant is the process owner. The process
owner has the permission to create/modify the owned
process model and to read all parent and derived pro-
cesses. After every change in a process model (e.g.,
in the main process model MPM), all directly derived
variants (V1, V2, and V3) are searched for an ele-
ment that is linked to the concerned element. If such
an element exists, the process owner of the variant
is notified about the change and can either accept or
decline it. If the change is accepted, then the next hi-
erarchy level is investigated and further sub-variants
(V1.1, V1.2, V1.3 as well as V2.1, V2.2) are consid-
ered. The change is forwarded until either a process
owner declines the change, no corresponding element
is found, or the leaves of the hierarchy are reached.
Finally, AdaVM implements rule-based graph
transformation techniques to automatically generate
process variants. The main advantages of graph trans-
formation techniques are that context-specific vari-
ants are dynamically generated and that modifica-
tions are not only defined for one variant but used
to generate several process variants (i.e., only a few
graph transformation rules are required to produce a
large number of process variants). In AdaVM, the al-
gebraic graph transformation approach described in
Section 2.1 is applied. The algebraic approach sup-
ports the above defined change patterns: an element is
inserted if it is only part of R (right-hand side), an el-
ement is deleted if it is only part of L (left-hand side),
and an element is modified if it is part of L and R but
the values of one or more attributes differ. A graph
transformation system can then either support vari-
ability by restriction (configurable approach) or vari-
ability by restriction and by extension (adaptable ap-
proach). A comparison of two graph transformation
systems implementing these techniques is provided in
the following section.
4 GRAPH TRANSFORMATION
In this section, the concepts of AdaVM are discussed
within a case study. The case study comprises a busi-
ness process for car services with various variants
for different engine types (i.e., fuel, diesel, electro)
and a distinction between basic (i.e., small) and ex-
tended (i.e., large) services. All variants are gener-
ated by applying algebraic graph transformation. An
advantage of graph transformation is the possibility
to deal with the exponentially increasing number of
variants with a linearly increasing number of graph
transformation rules, e.g., four activities each with
four variations may lead to 256 variants, which can
either be defined manually or automatically gener-
ated by 16 graph transformation rules (in ideal case).
To compare the configurable with the adaptable ap-
proach, two graph transformation systems have been
implemented in the tool AGG. The first graph trans-
formation system is called RM GT S, since the source
graph defines a comprehensive reference model (RM)
that comprises all possible process flows. The sec-
ond graph transformation system is called V M GT S,
since the source graph comprises the main (or default)
variant model (V M). For demonstration and evalua-
tion purposes, the two graph transformation systems
implement a subset of the AdaVM concepts (i.e., they
Modelling Business Process Variants using Graph Transformation Rules
69
Figure 3: Type Graph of RM GT S.
address the first and second requirement without links
and blocking). The full approach is currently imple-
mented in the BPM suite FireStart. Both graph trans-
formation systems prove the feasibility of the sug-
gested concepts but show differences regarding the
complexity.
The type graph of RM GT S is shown in Figure 3.
RM GT S requires abstract nodes and inheritance re-
lations within the type graph, so an attributed type
graph with inheritance is defined. The type graph
provides four abstract node types (Node, Activity,
Gateway, and Event) as highlighted by curly brack-
ets and italic font. Derived from node type Activ-
ity are all possible tasks of a car service (e.g., Oil-
change, Oil-filter-change, etc.). Abstract node type
Event is the parent of two concrete node types for
Start and End events, and node type Sequence is used
as dummy’ element to remove several subsequent
activities. Derived from exclusive gateways without
condition (ExclGate) and with condition (ExclGate-
Cond) are variant-specific exclusive gateways check-
ing whether diesel, electro, or large service is desired
(i.e., configurable elements). One further node type is
called Variables and used to store information regard-
ing the desired variant including attributes for engine
type and whether it should be a small or large service.
Finally, the type graph provides an edge label called
SF for general sequence flows and two further edge
labels SF Yes and SF No applied only after splitting
exclusive gateways. The type graph of V M GT S is
similar, but node types Sequence, ExclGate isDiesel,
ExclGate isElectro, and ExclGate isLarge are not re-
quired and omitted.
An example source graph of RM GT S is shown in
Figure 4. The process flows of all possible variants
are provided within one reference model and variant-
specific gateways are used to specify differences. The
desired variant is defined by element Variables (vari-
ant: large service for car with electro engine).
For transformation of a source to a target graph,
30 graph transformation rules with corresponding ap-
plication conditions have been defined in RM GT S.
Six of these rules are general and consider conditional
exclusive gateways. In addition, eight concrete rules
have been defined for each variant-specific gateway
(diesel, electro, large). The aim of these transforma-
tion rules is to generate the desired variant and to re-
move variant-specific gateways. The rule in Figure 5,
e.g., realises a large service by taking over the activity
from the Yes-Path and by removing the No-Path and
the surrounding gateways. A positive application con-
dition ensures that the rule is only applied if a variant
with a large service should be generated.
The resulting target graph after having ap-
plied all possible transformation rules is shown in
Figure 6. The transformation has replaced all
fuel- and diesel-specific activities with the activity
Check accumulator for electro engines. In addition,
activity Gear-maintenance for a large service was
included in the variant. Note that the other exclu-
sive gateway for large service was removed, since
Toothed-belt-maintenance and Air-filter-maintenance
are only relevant for fuel and diesel engines. In sum,
six variants can be automatically generated by chang-
ing the attribute values in the Variables element.
The second graph transformation starts with a de-
fault variant model and uses variability by restric-
tion and by extension to automatically generate fur-
MODELSWARD 2016 - 4th International Conference on Model-Driven Engineering and Software Development
70
Figure 4: Source Graph of RM GT S.
ther variants. An example source graph of V M GT S
is shown in Figure 7 and describes the default variant
for fuel engine and small service. An advantage of the
adaptable approach is that the source graph may al-
ready provide an executable variant. The Variables el-
ement again defines the desired variant (variant: large
service for car with electro engine).
For transformation of a source graph to all pos-
sible variants, only four graph transformation rules
with application conditions have been defined in
V M GT S. Each transformation rule actually corre-
sponds to one variant-specific gateway in the refer-
ence model of RM GT S. The graph transformation
rule shown in Figure 8, for example, inserts activity
Gear-maintenance in case of a large service (PAC) if
it has not yet been inserted (NAC). The application of
all possible transformation rules results in the same
target graph as shown in Figure 6.
The presented graph transformations are restricted
to structured diagrams (i.e., gateways are defined in
a bracket structure) and a basic set of process mod-
elling elements. However, further modelling elements
like sub-processes, additional events, or gateways are
only relevant for the business process and may appear
in process variants, but their influence on variant gen-
eration is limited. For a more detailed discussion on
configurable nodes and their influence see (Reichert
and Weber, 2012, p. 96ff).
According to Varr
´
o et al., the most important cor-
rectness properties of a trusted model transforma-
tion are termination, uniqueness (confluence), and be-
haviour preservation (Varr
´
o et al., 2006). Behaviour
preservation is not the aim of RM GT S and V M GT S,
since different variants with different behaviour are
generated. Thus, validation of the proposed graph
transformation systems is accomplished by proving
confluence and termination. These two properties en-
sure that the resulting graph is always the same, in-
dependent of the order of applied rules, and that rules
are not cyclic.
In order to prove confluence, it is either necessary
to show that each rule pair is parallel independent for
all possible matches or, in case of parallel dependent
rules, that the GTS is terminating and locally conflu-
ent. A GTS is locally confluent if all its critical pairs
(parallel dependent and minimal) are strictly conflu-
ent (Ehrig et al., 2006, p. 59ff, 144). RM GT S is
terminating, since all transformation rules delete el-
ements like gateways or activities. Only node type
Sequence is inserted by some rules but subsequently
removed by general rules. The full termination proof
comprising calculated rule, creation, and deletion lay-
ers is not presented due to space limitations. Further-
more, all critical pairs of RM GT S are of the same
rule and same match and, thus, isomorphic. Hence,
RM GT S is locally confluent and terminating, so the
whole graph transformation is confluent. The criti-
cal pairs of V M
GT S comprise different rules, but
strict AC-confluence could be proven for each pair.
A more challenging task was to prove termination of
V M GT S, since all transformation rules insert new
Modelling Business Process Variants using Graph Transformation Rules
71
(a) PAC (b) LHS (c) RHS
Figure 5: Transformation Rule of RM GT S.
Figure 6: Target Graph [Variant: Electro, Large Service].
node and edge types. Termination is ensured by neg-
ative application conditions (NAC: NotYetInserted).
However, since sequence flows are deleted in every
rule, all rules are automatically classified as deletion
layers, thereby ignoring the NACs. Thus, we pro-
pose to extend the deletion layer conditions presented
in (Ehrig et al., 2006, p. 31) to also consider NACs.
We further flattened the hierarchy of V M GT S with
the flattening algorithm presented in (Natschl
¨
ager and
Schewe, 2012) to 598 rules without inheritance rela-
tionships and proved termination for all node types.
Summing up, both graph transformation systems are
confluent and, thus, globally deterministic.
We then compared the graph transformation sys-
tems RM
GT S and V M GT S regarding their com-
plexity. Complexity is assessed by the number of
nodes in the type and source graphs as well as by the
amount of transformation rules and application con-
ditions. The complexity of RM GT S and V M GT S
is shown in Figure 9. Interestingly, V M GT S is less
Figure 7: Source Graph of VM GT S.
(a) PAC (b) NAC
(c) LHS
(d) RHS
Figure 8: Transformation Rule of V M GT S.
extensive in all measured aspects and, thus, also less
complex. The reason is that the rules of V M GT S
are more specific and explicitly define concrete ac-
tivities, e.g. Gear-maintenance, whereas the rules of
RM GT S use abstract node type Activity to represent
MODELSWARD 2016 - 4th International Conference on Model-Driven Engineering and Software Development
72
all possible concrete activities. Thus, an extension
of the process with a new variation would require an
adaptation of (every) reference model in RM GT S. In
contrast, V M GT S would either require an extension
of the default variant or (more likely) a new/extended
transformation rule. So, due to the differences in re-
quired adaptations, assessing the complexity of future
extensions is difficult. However, it can be assumed
that further extensions mainly increase the complex-
ity of source graphs of RM GT S and the amount of
rules and application conditions of V M GT S.
Figure 9: Complexity of RM GT S and V M GT S.
Summing up, we showed that graph transforma-
tion techniques are well-suited for variant manage-
ment. We, therefore, implemented two graph trans-
formation systems for generating process variants;
one supporting only variability by restriction, the
other supporting variability by restriction and by ex-
tension. Both graph transformation systems were
proven to be globally deterministic. Interestingly, the
adaptable approach was easier to implement (as it is
more specific) and it is less complex regarding the
type graph, source graphs, and the amount of rules
and application conditions. Thus, we recommend to
apply both techniques, variability by restriction and
by extension, for generating process variants. Other
concepts of AdaVM (e.g., linking, propagation of
changes, and visual highlighting of differences) do
not require verification and will be directly imple-
mented in BPM suite FireStart.
5 CONCLUSION
In this paper, we presented our concept for Adap-
tive Variant Modelling (AdaVM) as part of the Ada-
BPM framework for advanced adaptivity and excep-
tion handling in formal business process models. The
suggested approach requires fewer modifications of
the underlying notation, i.e., no configurable elements
or adjustment points are required, only an individ-
ual blocking of elements is proposed. In addition,
AdaVM considers linking of elements, propagation
of changes, and visual highlighting of differences in
process variants.
We further demonstrated that graph transforma-
tion techniques are well-suited for process variant
management and that variants can be automatically
created by a few graph transformation rules speci-
fying concrete variations. For that purpose, we pro-
vided a case study implementing two graph transfor-
mation systems; one supports variability by restric-
tion, the other supports variability by restriction and
by extension. The first system follows the config-
urable approach by defining the source graph as an
extensive basic model, which comprises all possible
process flows and can be configured to create vari-
ants. The second system follows the adaptable ap-
proach and the source graph comprises the main (de-
fault) variant model, which can be adapted by struc-
tural model adaptations to obtain new variants. We
proved both systems to be globally deterministic but
encountered differences regarding the complexity of
the graph transformation systems. Interestingly, the
case study reveals that the adaptable approach is less
complex regarding the type graph, source graphs, and
the number of rules and application conditions. Thus,
we intend to apply both techniques, variability by re-
striction and by extension, in AdaVM. Nevertheless,
our next step is to extend the case study to elaborately
evaluate the complexity of configurable and adaptable
approaches.
Summing up, the key differences of the proposed
concept for variability modelling, compared to other
state-of-the-art approaches, are (i) the support of vari-
ability by restriction and by extension with graph
transformation techniques, (ii) linking and propa-
gation of changes, (iii) individual blocking of ele-
ments/attributes, and (iv) visual highlighting of dif-
ferences in process variants. We, thus, expect that
the variant management implemented by BPM suite
FireStart outperforms other applications with respect
to functionality and usability.
REFERENCES
Aalst, W. v. d. (2013). Business process management: A
comprehensive survey. ISRN Software Engineering,
2013:1–37.
Aeneis (2014). aeneis Handbuch. intellior AG, aeneis 5.7
edition.
Axon IVY (2015). Axon.ivy 5.1 - Designer Guide. ivyTeam
AG.
Bizagi (2014). Bizagi BPM Suite - User Guide. Bizagi.
Doehring, M., Zimmermann, B., and Karg, L. (2011). Flex-
ible workflows at design- and runtime using BPMN2
Modelling Business Process Variants using Graph Transformation Rules
73
adaptation patterns. In Abramowicz, W., editor, Busi-
ness Information Systems, volume 87 of LNBIP, pages
25–36. Springer Berlin Heidelberg.
Ehrig, H., Ehrig, K., Prange, U., and Taentzer, G. (2006).
Fundamentals of Algebraic Graph Transformation.
Springer.
Gottschalk, F. (2009). Configurable Process Models. PhD
thesis, Technische Universiteit Eindhoven.
Hallerbach, A., Bauer, T., and Reichert, M. (2009). Guar-
anteeing soundness of configurable process variants in
provop. In IEEE Conference on Commerce and Enter-
prise Computing (CEC), pages 98–105.
Hallerbach, A., Bauer, T., and Reichert, M. (2010). Captur-
ing variability in business process models: The provop
approach. Journal of Software Maintenance and Evo-
lution: Research and Practice, 22:519–546.
Natschl
¨
ager, C., Kossak, F., and Schewe, K.-D. (2015). De-
ontic BPMN: a powerful extension of BPMN with a
trusted model transformation. Software & Systems
Modeling, 14(2):765–793.
Natschl
¨
ager, C. and Schewe, K.-D. (2012). A flattening ap-
proach for attributed type graphs with inheritance in
algebraic graph transformation. Electronic Communi-
cations of the EASST, 47.
Prologics (2014). FireStart Tutorial - Modellierung &
Ausf
¨
uhrung. Prologics.
Recker, J., Rosemann, M., van der Aalst, W., Jansen-
Vullers, M., and Dreiling, A. (2006). Configurable
reference modeling languages. In Fettke, P. and Loos,
P., editors, Reference Modeling for Business Systems
Analysis, pages 22–46. IGI Global, Pennsylvania.
Reichert, M. and Weber, B. (2012). Enabling Flexibility
in Process-Aware Information Systems: Challenges,
Methods, Technologies. Springer-Verlag Berlin Hei-
delberg.
Rosa, M. L., van der Aalst, W., Dumas, M., and Milani,
F. (2013). Business process variability modeling: A
survey. Technical report, Queensland University of
Technology.
Scheer (2015). Scheer BPaaS. Scheer.
Signavio (2015). Benutzerhandbuch. Signavio.
Varr
´
o, D., Varr
´
o-Gyapay, S., Ehrig, H., Prange, U., and
Taentzer, G. (2006). Termination analysis of model
transformations by Petri nets. In Corradini, A., Ehrig,
H., Montanari, U., Ribeiro, L., and Rozenberg, G., ed-
itors, Graph Transformations, volume 4178 of Lecture
Notes in Computer Science, pages 260–274. Springer
Berlin / Heidelberg.
Weber, B., Reichert, M., and Rinderle-Ma, S. (2008).
Change patterns and change support features - enhanc-
ing flexibility in process-aware information systems.
Data and Knowledge Engineering, 66(3):438–466.
MODELSWARD 2016 - 4th International Conference on Model-Driven Engineering and Software Development
74