The current issue and full text archive of this journal is available on Emerald Insight at:

www.emeraldinsight.com/1463-7154.htm

Process mining through artificial ANN and

support vector
neural networks and support machines
vector machines
1391
A systematic literature review
Received 6 February 2015
Ana Rocío Cárdenas Maita and Lucas Corrêa Martins Revised 20 April 2015
School of Arts, Sciences and Humanities, University of São Paulo, 24 May 2015
Accepted 9 June 2015
São Paulo, Brazil
Carlos Ramón López Paz
Faculty of Computer Engineering,
Higher Polytechnic Institute “José Antonio Echeverría”, Havana, Cuba, and
Sarajane Marques Peres and Marcelo Fantinato
School of Arts, Sciences and Humanities, University of São Paulo,
São Paulo, Brazil

Abstract
Purpose – Process mining is a research area used to discover, monitor and improve real business
processes by extracting knowledge from event logs available in process-aware information systems.
The purpose of this paper is to evaluate the application of artificial neural networks (ANNs)
and support vector machines (SVMs) in data mining tasks in the process mining context. The goal was
to understand how these computational intelligence techniques are currently being applied in
process mining.
Design/methodology/approach – The authors conducted a systematic literature review with three
research questions formulated to evaluate the use of ANNs and SVMs in process mining.
Findings – The authors identified 11 papers as primary studies according to the criteria established in
the review protocol. Most of them deal with process mining enhancement, mainly using ANNs.
Regarding the data mining task, the authors identified three types of tasks used: categorical prediction
(or classification); numeric prediction, considering the “regression” type, and clustering analysis.
Originality/value – Although there is scientific interest in process mining, little attention has been
specifically given to ANNs and SVM. This scenario does not reflect the general context of data mining,
where these two techniques are widely used. This low use may be possibly due to a relative lack of
knowledge about their potential for this type of problem, which the authors seek to reverse with the
completion of this study.
Keywords Process mining, Data mining, Artificial neural networks, Computational intelligence,
Support-vector machines
Paper type Literature review

1. Introduction
Business process management (BPM) includes methods, techniques and tools to
support the design, enactment, management and analysis of business processes,
previously called business workflows, or simply workflows (van der Aalst et al., 2003).
The BPM life cycle includes the following stages: business processes modeling; Business Process Management
implementation of business process models; execution and supervision of business Journal
Vol. 21 No. 6, 2015
pp. 1391-1415
This work was supported by Fapesp (São Paulo Research Foundation) and Capes (Coordination © Emerald Group Publishing Limited
1463-7154
for the Improvement of Higher Education Personnel), Brazil. DOI 10.1108/BPMJ-02-2015-0017
BPMJ process instances; monitoring and auditing of running business process instances; and
21,6 evaluation and improvement of business process models (Weske, 2012). In this last
stage, the execution and monitoring histories of the instances of a business process can
be evaluated aiming its optimization.
According to Han and Kamber (2006), data mining refers to extracting or mining
knowledge from lage amounts of data. Moreover, data mining can be seen as a stage in
1392 a larger process called Knowledge Discovery from Data (Fayyad et al., 1996). In this
larger process, data from a specific context is collected, transformed and organized
before undergoing mining. After mining, the result should be organized in a structure
that is accessible to direct human interpretation. There are different tasks to be
addressed in data mining, with different techniques that can be applied to each one.
Among these techniques, those from the computational intelligence area are commonly
used (Wang and Fu, 2005), especially those using inductive reasoning, such as artificial
neural networks (ANN) and support vector machines (SVM), since they can be applied
in different data mining tasks (Stahl and Jordanov, 2012; Cortez, 2010). SVM, for
example, was included among the top ten algorithms in data mining according to the
sixth IEEE International Conference on Data Mining held in 2006 (Wu et al., 2007). In
addition, both techniques are quite capable of conducting non-complex linear
mappings, common in data mining contexts, and therefore the potential to achieve
highly accurate predictions. Finally, these models can extract knowledge, which is
useful in decision-making environments (Cortez et al., 2009).
The combination of both areas – BPM and data mining – establishes a new field of
study, known as business process mining, or simply process mining (van der Aalst,
2011). It deals with applying data mining tasks on data from the BPM lifecycle. Its goal
is to extract knowledge about events/data from the work carried out in the different
phases of a business process, seeking to improve it, discovering associations between
variables, behavior or misbehavior patterns (van der Aalst, 2011).
In this context, the main objective of this paper is to present the results of a study
that identified and analyzed the primary studies related to process mining that
exclusively use ANN or SVM as the data mining technique. Our intention was to reach
an overview of this type of process mining approach and also to more carefully
explore the use of two of the most important computational intelligence techniques.
With this purpose, this paper presents the following sections: a theoretical background;
the detailed SLR methodology; the results overview; the analysis of the selected
primary studies; a discussion on the results; the validity threats analysis; and, finally,
the conclusions.

2. Theoretical background
This section presents the main concepts related to this SLR: data mining and
process mining.

2.1 Data mining

Overall organizations have been effective in extracting, organizing and storing data.
However, transforming this large amount of data into useful knowledge is still difficult.
In this context, data mining is used to find useful knowledge, often unexpected, from
the analysis of the data produced in various types of systems and processes, through
the disclosure of behavior patterns that can serve as a basis for high-strategic decision
making (Nazem and Shin, 2002).
 Accordingly, data mining explores and analyzes data, automatically or semi- ANN and
automatically, in order to discover relevant patterns, hidden to the database. The main support vector
data mining tasks are: categorical prediction (or classification); numeric prediction
(“regression” type); numeric prediction (“trend identification” type); clustering analysis;
machines
and, frequent discovery patterns, discovering association and correlation rules (Han
and Kamber, 2006). These tasks can predict trends and behaviors, enabling to make
decisions based on accumulated knowledge, and often overlooked, contained in the 1393
databases themselves. Thus, during the data mining tasks, existing but unknown
patterns and relationships can be identified, leading to generating hypotheses that are
useful to the user.
For each type of data mining task, different techniques can be applied (Han and
Kamber, 2006; Wu et al., 2007), including those in the computational intelligence area
(Wang and Fu, 2005). Computational intelligence techniques have proven to be efficient
in solving data mining tasks because of their good ability to handle data from ill-
defined contexts, where the rigid and highly accurate and restricted models are
impractical, something often found on data from real-world problems (Abonyí et al.,
2005). Furthermore, computational intelligence techniques that deal with approximate
reasoning (such as fuzzy logic and probabilistic reasoning), inductive reasoning (such
as ANN and SVM) and evolutionary reasoning (such as evolutionary computing
techniques) are highly suitable for data mining since its context involves empirical and
behavior knowledge, in which information is inaccurate, incomplete and imperfect
(Abonyí et al., 2005).
Regarding the techniques that use inductive reasoning – ANN and SVM – they are
both similar, in how to explore a data mining problem and also in how to provide the
solution, regarded as equivalent in their most basic forms (Cristianini and Shawe-
Taylor, 2000). Nevertheless, SVM has additional characteristics in the optimization
solution model, since it also has the structural risk minimization of the model, besides
the empirical risk minimization (Vapnik, 1995). ANN, in its several architectures, and
SVM, in its several variations, represent together a powerful set of tools widely used in
solving prediction and clustering analysis tasks for data mining.

2.2 Process mining

Process mining (also known as workflow mining) is a relatively new research area that
lies between data mining and machine learning, on the one hand, and business
processes modeling and analysis, on the other hand (van der Aalst, 2011). Its objective
is to discover, monitor and improve real business processes through knowledge
extraction from event logs readily available in process-aware information systems
(van der Aalst, 2011).
Many process mining approaches currently use heuristic techniques and
algorithms; that is, “golden rules” based on assumptions about business process
patterns (van der Aalst, 2011). However, there are also some process mining
approaches that use computational intelligence techniques, including evolutionary
computing and ANN (Tiwari and Turner, 2008). Nevertheless, these computational
intelligence techniques are still in the minority in the process mining context when
compared to heuristics techniques and other techniques such as statistical techniques.
Although ANN is one of the most widely used techniques in data mining, the SLR
presented by Tiwari and Turner (2008), which provides a comparison of key research
efforts relating to process mining problems and techniques, with more than 50 papers,
list only one study using ANN.
BPMJ 3. Methodology
21,6 A systematic literature review (SLR) is a means of identifying, evaluating, and
interpreting all available research relevant to a particular research question, or topic
area, or phenomenon of interest (Kitchenham, 2007). Individual studies contributing to
a SLR are called primary studies; a SLR is a form of secondary study. A SLR differs
from traditional reviews (such as simple literature reviews) and surveys with
1394 comments made by domain experts because the experts use a replicable, scientific, and
transparent approach to avoid bias (Biolchini et al., 2005).
To conduct this SLR, we used the guidelines by Kitchenham (2007). According to
these guidelines, a SLR includes several steps, which can be grouped into three main
phases: planning the review; conducting the review; and reporting the review. Sections
3.1 and 3.2 describe, respectively, details of the review planning and of the review
conduction. Section 4 presents the review results.

3.1 Planning the SLR

The planning of the SLR includes the following steps: (i) identifying the need for a
review; (ii) commissioning a review; (iii) specifying the research questions; fourth,
developing a review protocol; and finally, evaluating the review protocol. The following
subsections details steps (i), since they are considered the main and mandatory steps of
the review planning.
3.1.1 Need for a SLR. The authors of this paper are part of an interdisciplinary
research group in the area of information systems, which develops work in the data
mining areas, mainly using computational intelligence and BPM techniques. Therefore
they are interested in conducting work in the area of process mining, using
computational intelligence techniques and thereby understand and evaluate the
primary studies published within that context.
Considering that thus far no SLR has been published for this purpose, these authors
deemed important and useful undertaking such a review and consequently sharing it
with other potentially interested researchers. This way other research groups can start
their work from this review, saving research time and resources to focus on proposals
and evaluations of new approaches, using the existing works as a basis. We chose the
SLR for this study because of its methodological benefits, as it maximizes the
possibility of recovering complete data sets, minimizing bias possibility.
3.1.2 Research questions. Assuming the need for a SLR, we defined a set of specific
research questions for this work. These questions are mostly directed to discovering
how certain computational intelligence techniques are applied in the process mining
area, aiming at a comprehensive understanding of this joint research area. The main
purpose of this study is to identify and assess such works that intend to use ANN or
SVM in the context of process mining:
RQ1. What types of ANN or SVM are used to implement data mining tasks when
specifically applied to the process mining context?
For this question, we considered the following types of ANN: first, in terms of the type
of architecture, which refers to how their processing units (neurons) are structured, the
networks can be classified as feedforward single-layer, recurrent single-layer,
feedforward multilayer, recurrent multilayer, and competitive layer (Haykin, 2005);
and second, with regards to the type of learning, which refers to the strategy used by
the algorithm to correct or approximate the mapping model or the decision surface that
are being sought, the networks can be classified as supervised learning, unsupervised ANN and
learning, reinforcement learning and fixed-weight networks (Fausett, 1994). In terms of support vector
the types of SVM, we consider the basic model SVM for classification (with supervised
learning) and its two main variations support vector regression (SVR) (with supervised
machines
learning) and support vector clustering (SVC) (with unsupervised learning) (Haykin,
2005; Cristianini and Shawe-Taylor, 2000):
RQ2. Which data mining tasks are addressed in process mining when applying
1395
ANN and SVM?
For this question, we considered the main data mining tasks (Han and Kamber, 2006):
first, categorical prediction (or classification) – which consists of finding a function that
maps (or classifies) a certain event (or fact) in a class (or category) from a finite and
predefined set of classes; second, numeric prediction (“regression” type) – which consists
of finding a function that approximately maps a particular event for a numeric value;
third, numeric prediction (“trend identification” type) – involves a time component, that
is, it represents that specific class of problems for which they want to predict the
behavior or future value of a particular variable of interest based on previous values of
this variable and possibly others; fourth, clustering analysis – consists of finding
similarities between data to define a finite set of clustering data, organizing the data as
clusters in order to maximize intra-cluster similarity and minimize intercluster similarity,
or also to determine an unknown set of classes for the data in order to also describe them;
and finally, frequent patterns discovery, association and correlation rules discovery –
which consists of identifying frequent events which may be directly or indirectly
associated (or correlated) with the occurrence of events or transactions:
RQ3. What types of process mining are treated by ANN and SVM when these
techniques are used to address data mining tasks in this context?
For this question, we considered the following types of process mining: first,
discovery – which refers to producing a still unknown business process model based
solely on an event log, using no prior information; second, conformance – refers to
comparing a business process model already known as an event log of the same
business process in order to check whether the reality, as recorded in the log, is in line
with the business process model and vice versa; and finally, enhancement – refers to
modifying an existing business process model based on an event log of the same
business process in order to, for example, repair a process model so that it will better
achieve its purpose (van der Aalst, 2011).
3.1.3 SLR protocol. According to Kitchenham (2007), a SLR protocol specifies the
methods to be used to undertake a specific review. Moreover, a pre-defined protocol
is essential to reduce the possibility of research bias; it describes how to search and
select relevant primary studies and analyze the data extracted to answer the
predefined research questions. The main components of a SLR protocol include the
definition of data sources; search strategy; primary study selection strategy;
extraction method; and data summary.
A. Data sources and search strategy. We chose the following data sources to search
for works in this SLR: Scopus[1] and ISI Web of Science (WoS)[2]. These data sources
were used in order to maximize the number of candidate works found, since together
they index most of the existing papers in the most important digital libraries, such as:
IEEEXplore, ACM Digital Library and SpringerLink.
BPMJ For the search strategy, the search string shown in Table I was built using
21,6 keywords derived from the aforementioned research questions, connected through the
logical connectives AND and OR. We elaborated this search string seeking to maximize
the number of candidate works to be found for this SLR. Therefore we sought to use a
broad range of options (among words, expressions, and acronyms). First, the search
string seeks to find works that are generically related to process mining; and second,
1396 the string seeks to limit this search to works that specifically deal with some aspect
related to ANN, including its main architectures, or SVM, including its two main
variants. The keywords used in the string were defined based on the knowledge and
information of one of the authors of this study, an expert on ANN and SVM.
Since it was not possible to use the same exact string for both data sources, we had
to adjust the base search string shown in Table I so that it could be directly applied to
the search engines of each one of the data sources used. Table II presents the search
strings modified to fit the characteristics of each one of these engines. These adjusted
search strings have already embedded some additional restrictions, as part of this SLR

(“process mining” OR “workflow mining” OR “process mining” OR “mining workflow” OR ((“business

process” OR workflow) AND “data mining”)) AND ((neural OR SOM OR “Self Organizing” OR “Self-
Organizing” OR “Organizing Map” OR MLP OR “Multilayer Perceptron” OR “Backpropagation” OR
“Back-Propagation” OR RBF OR “Radial Basis Function” OR artmap OR “Adaptive Resonance
Table I. Theory” OR hopfield OR LVQ OR “Learning Vector Quantization”) OR (“SVM” OR “SVR” OR “SVC”
Search string OR “Support Vector” OR “Support-Vector”))

Scopus
TITLE-ABS-KEY((“process mining” OR “processes mining” OR “workflow mining” OR “workflows
mining” OR “mining process” OR “mining processes” OR “mining workflow” OR “mining workflows”
OR ((“business process” OR “business processes” OR workflow OR workflows) AND “data mining”))
AND ((neural OR SOM OR “Self Organizing” OR “Self-Organizing” OR “Organizing Map” OR
“Organizing Maps” OR MLP OR “Multilayer Perceptron” OR “Backpropagation” OR “Back-
Propagation” OR RBF OR “Radial Basis Function” OR artmap OR “Adaptive Resonance Theory” OR
hopfield OR LVQ OR “Learning Vector Quantization”) OR (SVM OR SVR OR SVC OR “Support
Vector” OR “Support-Vector”))) AND PUBYEAR W 2003 AND (LIMIT-TO(LANGUAGE, “English”))
AND (LIMIT-TO(DOCTYPE, “cp”) OR LIMIT-TO(DOCTYPE, “ar”) OR LIMIT-TO(DOCTYPE, “ip”))
AND (LIMIT-TO(SUBJAREA, “COMP”) OR LIMIT-TO(SUBJAREA, “ENGI”) OR LIMIT-TO
(SUBJAREA, “MATH”) OR LIMIT-TO(SUBJAREA, “BUSI”) OR LIMIT-TO(SUBJAREA, “MULT”))
ISI Web of Science
Topic ¼ ((“process mining” OR “processes mining” OR “workflow mining” OR “workflows mining” OR
“mining process” OR “mining processes” OR “mining workflow” OR “mining workflows” OR
((“business process” OR “business processes” OR workflow OR workflows) AND “data mining”)) AND
((neural OR SOM OR “Self Organizing” OR “Self-Organizing” OR “Organizing Map” OR “Organizing
Maps” OR MLP OR “Multilayer Perceptron” OR “Backpropagation” OR “Back-Propagation” OR RBF
OR “Radial Basis Function” OR ARTMAP OR “Adaptive Resonance Theory” OR Hopfield OR LVQ OR
“Learning Vector Quantization”) OR (“SVM” OR “SVR” OR “SVC” OR “Support Vector” OR “Support-
Vector”))); Refined by: Language: (ENGLISH) AND Document Types: (PROCEEDINGS PAPER OR
PAPER) AND Research Ares: (COMPUTER SCIENCE OR ENGINEERING OR OPERATIONS
Table II. RESEARCH MANAGEMENT SCIENCE OR MATHEMATICAL COMPUTATIONAL BIOLOGY OR
Search strings MATHEMATICS OR MEDICAL INFORMATICS OR BUSINESS ECONOMICS OR AUTOMATION
specific for each CONTROL SYSTEMS OR SCIENCE TECHNOLOGY OTHER TOPICS) Timespan ¼ 2004-2014.
data source Databases ¼ SCIEXPANDED, CPCI-S, BKCI-S
protocol (as explained in the following inclusion and exclusion criteria), when the ANN and
specific search engines for each data source enable doing it, in order to perform a pre- support vector
filter to discard applicable records for this SLR.
We first applied the search at the end of 2013 and reapplied it at the end of 2014.
machines
B. Primary study selection strategy. We specified a set of inclusion and exclusion
criteria based on the analysis scope and the quality of the papers found to guarantee
that only works actually related to the context of process mining with ANN or SVM 1397
would be selected as the primary studies. A study, identified after applying the search
string, is selected as primary study if it meets all the predefined inclusion criteria; but it
is eliminated if it meets any of the predefined exclusion criteria. Considering the
purpose of this SLR, as well as its respective identified research questions, we defined
the following criteria:
B.1. Inclusion criteria
IC-1 The paper essentially addresses data mining, i.e., data mining is directly related
to the main scope of the work rather than to data mining terms merely mentioned in a
generalized manner.
IC-2 The paper essentially addresses business processes (including workflow),
i.e., business processes and/or workflow rely directly to the main scope of the work rather
than to business processes/workflow terms merely mentioned in a generalized manner.
IC-3 Both data mining and business processes are addressed together in the paper to
present a process mining approach rather than each one of them being addressed
independently.
IC-4 The paper presents as its main objective the use of ANN or SVM as a technique
to implement data mining tasks in the process mining context.
B.2. Exclusion criteria
EC-1 The paper is not electronically available on the web.
EC-2 The paper is not presented entirely in the English language.
EC-3 The paper is not related primarily to the Computer Science or Information
Systems fields (e.g. the paper is related primarily to Medicine or Industrial
Engineering).
EC-4 The data register identified after applying the search string does not actually
refer to a scientific paper, but to some non-peer reviewed publication, such as: technical
reports; books and book chapters; proceedings’ prefaces; and journal’s editorials.
EC-5 The paper presents some type of review, such as a survey or some SLR
(i.e. a secondary study), and not the outcomes of some specific research work
(i.e. a primary study).
EC-6 The paper was published over 11 years ago.

3.2 Conducting the SLR

This section presents the main steps to conduct the SLR. They are grouped in
two main steps identification and selection of primary studies as well as the quality
assessment of the primary studies; and data extraction and synthesis of the primary studies.
3.2.1 Identification, selection and quality assessment of primary studies. We based the
identification and selection of primary studies on the strategy proposed by Chen and Ali
Babar (2011), which consists of three steps as shown in Figure 1 and described as follows:
• Step 1. Identifying relevant primary studies on search databases: to identify
candidate primary studies, we applied the specific search strings of both
BPMJ
21,6 Search Strategy

1. Identifying relevant studies on search

1398 databases
Result = 188

2. Applying the inclusion and exclusion criteria

Result = 11

Figure 1. 3. Assessing the quality of primary studies

Primary studies adapted
selection process Result = 11

data sources (see Table II) to the respective search engines. The searches
returned 188 registers, considering the sum of both data sources
and eliminating duplicate results. We regard this as a reasonable
number since we significantly delimited the search string considering the
given research area.
• Step 2. Applying the inclusion and exclusion criteria: in this step, we analyzed for
inclusion or exclusion, based on the predefined criteria, each data register
returned and identified as a candidate to be a primary study (as defined in
Section 3.1.3). We applied the inclusion and exclusion criteria based on a
thorough reading and analysis of the title, abstract and keywords of the
respective paper. To provide more reliability to the result, when it was not
completely clear if the paper should be included or excluded only with this first
reading, we also considered the introduction and conclusion sections. As a result
of applying the inclusion and exclusion criteria, out of the 188 initial candidate
registers, we selected only 11 primary studies, considered as the primary studies
for this SLR.
• Step 3. Assessing the quality of primary studies: this step is considered critical
and mandatory to assess the quality of primary studies (Kitchenham, 2007).
However, specifically for this SLR, no further exclusions were done based on the
quality assessment, as originally proposed by Chen and Ali Babar (2011).
Alternatively, we only carried out the quality assessment considering four
attributes and 11 sub-attributes used to categorize the primary studies through
an adaptation of the evaluation approach proposed by Dyba and Dingsøyr (2008).
Table III describes the attributes – “reporting”, “rigor”, “credibility” and
“relevance” and their sub-attributes.
Table IV shows a summary of the 11 primary studies, resulting from Step 2, described
above. The data show that seven were published in journals and four were published in
conferences.
Attribute Description
ANN and
support vector
Reporting Concerned with the quality of the information being reported in terms of a study’s rationale, machines
objectives, and context
Rep.1 Is the primary study based on research (or merely a “lessons learned” report based on
expert opinion)?
Rep.2 Is there a clear statement of the research goals?
Rep.3 Is there an adequate description of the context in which the research was carried out? 1399
Rigor Concerned with the accuracy of the research methods employed to establish the validity of data
collection tools and the analysis methods, and hence the trustworthiness of the findings
Rig.1 Was the research design appropriate to address the research goals?
Rig.2 Was the recruitment strategy appropriate for the research goals?
Rig.3 Was there a control group to compare the treatments?
Rig.4 Was the data collected in such a way that it addressed the research issue?
Rig.5 Was the data analysis sufficiently rigorous?
Credibility Concerned with the trustworthiness of the study methods used for ensuring that the findings
were valid and meaningful
Cre.1 Was the relationship between researcher and participants adequately considered?
Cre.2 Is there a clear statement of the findings?
Relevance Concerned with the assessment of the significance of the study for the software industry at
large and the research community Table III.
Rel.1 Is the study of value for research or practice? Quality assessment
Source: Based on Dyba and Dingsøyr (2008) attributes

3.2.2 Data extraction and synthesis. During this stage, all selected primary studies
underwent more in-depth readings, going beyond their introduction and conclusion
sections, mainly seeking to answer the research questions specified (as presented in
Section 3.1.2). Moreover, during such a detailed reading and analysis, the application of
exclusion criteria could be refined in some cases. Specifically for data extraction, we
used a form to capture relevant information from the primary studies, in order to
answer the research questions. We used a set of common attributes, in such a form, to
extract data from each primary study; these attributes were: publishing vehicle
(workshop, conference, or journal); title; and, author(s). Besides theses common
attributes, we also used a series of other specific attributes during the paper analysis
and the completion of the form; these specific attributes were used to extract data
taking into account each of the different research questions, to which specific
extraction strategies were also established.

4. Results overview
This section presents an overview of the results produced by conducting this SLR
according to the protocol presented in the Section 3. Figure 2 shows the temporal
distribution of the primary studies (from 2004 to 2014). According to the data
presented, there is a concentration of primary studies published in the last four years of
the period assessed.
An analysis of the selected primary studies produced the classifications and
structured results presented as follows, specifically related to the assessment of the
quality attributes and the responses of the research questions.
Quality assessment. Assessing the quality of primary studies.
BPMJ Publication
21,6 Year ID Authors Title type

2004 J-1 Ho, Lau, Ip and Ning An intelligent information infrastructure to Journal
support the streamlining of integrated logistics
workflow
2006 J-2 Ho, Lau, Lee, Ip and Pun An intelligent production workflow mining Journal
1400 system for continual quality enhancement
C-1 Tsang, Lau and Kwok Development of a distributed process mining Conference
system for reactive ion etching enhancement
C-2 Thomas, Adam, Leyking A fuzzy paradigm approach for business Conference
and Loos process intelligence
2011 J-3 Kang, Lee, Kim and Kang KNNI-SVM: A hybrid algorithm integrating Journal
imputation and support vector machine for
real-time business process monitoring
C-3 Zhang, Yang, Shen and Process optimization of candy production Conference
Chen based on data mining
2012 J-4 Liu, Cheng and Ni Mining event logs to support workflow Journal
resource allocation
2013 J-5 Song, Yang, Siadat and A comparative study of dimensionality Journal
Pechenizhiy reduction techniques to enhance trace
clustering performances
C-4 Esteva, Xu, Felix and Data mining for “big archives” analysis: A case Conference
Anantha study
2014 J-6 Kranjc, Smailovića, Active learning for sentiment analysis on data Journal
Podpečana, Grčara, streams: Methodology and workflow
Žnidaršiča, Lavrača implementation in the ClowdFlows platform
Table IV. J-7 Seret, Broucke, Baesens and A dynamic understanding of customer Journal
Final list of primary Vanthienen behavior processes based on clustering and
studies sequence mining

Number of studies
3

2
Figure 2.
Distribution of 1
primary studies by
publication year 0
2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014

Table V presents data regarding the quality evaluation of the selected primary studies
taking into account the attributes defined in Table III. Each sub-attribute was assessed
and scored as “Complete – 1.0 point”, “Partial – 0.5 point” or “None – 0 point”. Studies
were best assessed in terms of the attributes “reporting” and “relevance”, with relative
rates of 91 and 82 percent, respectively. In contrast, they were worst assessed in terms
of the attributes “credibility” and “rigor”, with relative rates of 61 and 56 percent,
respectively:
RQ1. What types of ANN or SVM are used to implement data mining tasks when
specifically applied to the process mining context?
 Reporting Rigor Credibility Relevance
ANN and
ID Rep.1 Rep.2 Rep.3 Rig.1 Rig.2 Rig.3 Rig.4 Rig.5 Cre.1 Cre.2 Rel.1 Quality score support vector
machines
J-1 C C C C P N P P P C C 8.0
J-2 C C P C C N C P P C C 8.5
C-1 C P P C N N C P N P P 5.5
C-2 C C P P N N P N N P P 4.5
J-3 C C C C N N P N N P P 5.5 1401
C-3 C P P P N N P N N P P 4.0
J-4 C C C C C N C C P C C 9.5
J-5 C C C C C N C C C C C 10.0
C-4 C C C C C N C C C C C 10.0
J-6 C C C C P N C P P C C 8.5 Table V.
J-7 C C C C C N C C P C C 9.5 Classification of
Notes: C, complete (1.0 point); P, partial (0.5 point); N, none (0 point) primary studies
Source: Dyba and Dingsøyr (2008) regarding quality

Tables VI, VII and VIII illustrate a summary of the data extraction and classification
result from the 11 primary studies scope of this SLR in order to answer this first
research question. With regard to the approaches identified in these primary studies
using ANN, the data of Table VI and Table VIII show that most of them have
multilayer architecture with supervised learning. As for the approaches identified in
these primary studies using SVM, the data of Tables VII and VIII show that all studies
used the basic model for SVM classification with supervised learning:
RQ2. Which data mining tasks are addressed in process mining when applying
ANN and SVM?

Feedforward single- Recurrent single- Feedforward Recurrent Competitive

ID layer layer multilayer multilayer layer

J-1 X
J-2 X
C-1 X
C-2 X
C-3 X Table VI.
J-5 X Classification of
J-7 X primary studies
Note: X means that the evaluated attributes (at the first line) were found in the evaluated primary regarding ANN
studies (at the first column) architectures

ID SVM for classification SVR SVC

J-3 X
J-4 X Table VII.
C-4 X Classification of
J-6 X primary studies
Note: X means that the evaluated attributes (at the first line) were found in the evaluated primary regarding SVM
studies (at the first column) variations
BPMJ In relation to data mining tasks, the data in Table IX show that for the approaches
21,6 identified in the primary studies scope of this SLR, three of the five categories are
covered when considering its resolution with ANN or SVM. Representatives for the
following tasks were found: categorical prediction (or classification), numeric prediction
(“regression” type), and clustering analysis. No study has been identified for the other
two categories: numeric prediction (“trend identification” type), and frequent patterns,
1402 association and correlation rules discovery. In fact, two studies involve frequent
patterns discovery task, but are solved with techniques other than ANN and SVM, thus
not counted in Table IX:
RQ3. What types of process mining are treated by ANN and SVM when these
techniques are used to address data mining tasks in this context?
With regards to the type of process mining, whose objective is to verify the ANN and
SVM applications in the area of interest, the data in Table X show that two studies

ID Supervised learning Unsupervised learning Reinforcement learning Fixed-weight networks

J-1 X
J-2 X
C-1 X
C-2 X
J-3 X
C-3 X
J-4 X
J-5 X
Table VIII. C-4 X
Classification of J-6 X
primary studies J-7 X
regarding ANN and Note: X means that the evaluated attributes (at the first line) were found in the evaluated primary
SVM learning type studies (at the first column)

Numeric prediction
Categorical “Trends frequent patterns discovery,
prediction (or “Regression” identification” Clustering association and correlation rules
ID classification) type type analysis discovery

J-1 X
J-2 X
C-1 X
C-2 X
J-3 X
C-3 X
J-4 X
J-5 X
Table IX. C-4 X
Classification of J-6 X
primary studies J-7 X
regarding data Note: X means that the evaluated attributes (at the first line) were found in the evaluated primary
mining task type studies (at the first column)
ID Discovery Conformance Enhancement
ANN and
support vector
J-1 X machines
J-2 X
C-1 X
C-2 X
J-3 X
C-3 X 1403
J-4 X
J-5 X
C-4 X Table X.
J-6 X Classification of
J-7 X primary studies
Note: X means that the evaluated attributes (at the first line) were found in the evaluated primary regarding the
studies (at the first column) process mining type

relate to the discovery type, three to enhancement, and six to conformance. Although
this classification was carried out considering the definitions given by van der Aalst
(2011), most studies in fact do not strictly follow the definitions proposed by that expert
for the process mining area. Therefore the classification in Table X refers to an
approximation of that proposal by van der Aalst (2011) taking into account the analysis
conducted by the authors of this study.

5. Primary studies analysis

This section presents an overview of the analysis undertaken on the 11 primary studies
selected as part of this SLR. First, we present the two studies related to the process
discovery type of process mining; then, we present the three studies related to the
process conformance type; and, finally, we present the six studies related to the process
enhancement type.

5.1 Process mining – discovery type

( J-5) A comparative study of dimensionality reduction techniques to enhance trace
clustering performances (2013).
This study examines the effects of applying dimensional reduction techniques on
data sets to be subjected to clustering techniques. Dimension reduction is commonly
applied to mining problems that analyze data sets with a high number of descriptive
characteristics, that is, which are located in high-dimensional spaces. In this work, the
trace clustering task was used to improve the understanding of process models
discovered from the analysis of log events. The authors used real event logs recorded
from patient treatment processes in a Dutch hospital, on which algorithms were applied
for discovering process models. Three popular dimensional reduction techniques were
applied to the data: singular value decomposition; random projection; and principal
component analysis. Next, three clustering techniques were used to implement the trace
clustering: ANN self-organizing maps (SOM), k-means and hierarchical agglomerative
cluster. The experimental results achieved show that applying dimensional reduction
result in gains, in terms of the computational time spent on the subsequent clustering
task, and in some cases an improved result of the clusters formed, depending on the
combination of techniques used and the data set context. Specifically for ANN SOM,
improved clustering results were observed only for the context of simple data sets,
BPMJ regardless of the dimension reduction techniques used. As for the k-means, for
21,6 example, the dimension reduction techniques were essential for obtaining better
results. Although not discussed by the authors, the fact that ANN SOM did not benefit
from the prior dimensional reduction as much as k-means is expected, since ANN SOM
has the ability to make projections of high-dimensional spaces for low-dimensionality
spaces, which makes it easier to analyze data in higher dimensions, when compared to
1404 the k-means.
( J-7) A dynamic understanding of customer behavior processes based on clustering
and sequence mining (2014).
This work focusses on a two-step clustering based approach, incorporating both ANN
SOM and k-means to identify coordinate sequences used as input for a frequency
sequence mining technique. These methods are combined to discover prominent
customer behavior trajectories in databases that represent the “behavior process” for
particular groups of customers. Customer trajectories are represented as coordinate
sequences and are analyzed to discover the most frequent ones and therefore reveal the
organization’s most frequent client behaviors. During the clustering task, ANN SOM is
used to provide the customers’ trajectory design, over the relationship time with the
organization, between the different clusters representing the behavior patterns.
The data projection capability in two-dimensional maps allows discovering these
trajectories. The clusters are discovered by applying the k-means algorithm on the
two-dimensional projection created by ANN SOM. Thus, ANN SOM neurons, which
represent the customers (from descriptive characteristics), which are organized in
a two-dimensional topological structure, are used as input to form clusters by the
k-means. This approach can help explain customer actions influenced by consumer
decisions and improve business processes. It refers to a data-driven approach using
state sequences derived from instance level attributes collected over time (rather than
events logs, which are usually used in process mining), meaning that no full event log
data is required to use it.

5.2 Process mining – conformance type

( J-1) An intelligent information infrastructure to support the streamlining of integrated
logistics workflow (2004).
This paper presents an intelligent information system using ANN together with
OLAP[3] to simplify the information flow in a logistics process and identify
anomalous behavior when executing process instance. Simplifying the process is based
on the organization’s behavior learning through an ANN fed with data from an OLAP
system, which stores the entire information flow of the organization. From the
observation of the information flow of the instance process running, the ANN can
identify and report the risk of an anomalous behavior, which can be used to support
remedial actions. To illustrate the approach, the authors present a case study of an
airline company, where there is a potential risk of a logistics process problem involving
an aircraft or a part monitored by ANN powered by data from an OLAP system, which
in turn receives information about aircraft conditions, airport conditions, fuel types,
etc. This study shows benefits of combining the OLAP and ANN technologies,
suggesting that it could convert the raw data into knowledge with self-learning
capability to thereby minimize the time of the analytical process associated with
traditional methods.
( J-3) KNNI-SVM: a hybrid algorithm integrating imputation and SVM for real-time
business process monitoring (2011).
 This paper proposes an approach to monitor real-time business processes using an ANN and
approach that integrates K Nearest Neighbor Imputation (KNNI) and SVM. KNNI is support vector
used to impute to attributes, values not yet observed, in running process instance,
based on historical data, in other words, event logs. And SVM is used to classify these
machines
processes in two possible classes: success and error. For this classification, the SVM
classifier is trained using preprocessing that is also based on historical data,
considering cases with known end result in terms of success or failure of the process 1405
instance. Thus, the event logs are analyzed by SVM in order to find correlations
between the attributes and the final results, providing a decision model that can predict
success or error outcomes still during the monitoring of the process instance being
executed. The results of this work show that adopting an early warning system enables
efficient understanding of the state of the ongoing instance and helps industrial
managers by providing appropriate operations so that they can prevent failures rather
than wait until their actual occurrences.
( J-6) Active learning for sentiment analysis on data streams: methodology and
workflow implementation in the ClowdFlows platform (2014).
This work introduces a cloud-based scientific workflow platform, called
ClowdFlows, which is able to perform on-line dynamic adaptive sentiment analysis
of microblogging posts (Twitter posts). The authors developed a web service that uses
a SVM algorithm to build and update sentiment analysis models and minimize the
effort required to apply labels to tweets. For this purpose, the authors developed a web
page – called active learning workflow for sentiment analysis, in which the model was
periodically updated and applied to new instance. The main idea of the ClowdFlows
platform is the workflow model that consists of an abstract representation of
workflows and workflow components: the operation on the server is validated, and a
success or error message with additional information is passed to the user interface.
The platform has successfully passed the stress test, which means that it supports
many concurrent users that continuously execute workflows. The authors also
intended to show the adaptation of the ClowdFlows platform, enabling it to work on
real time data streams. As a result, workflows in ClowdFlows are no longer limited to
static data on the server but can connect to multiple data sources and can process the
data continuously. Although this work illustrates a reusable analysis model, which can
help monitoring Twitter sentiments in other areas, it does not represent a great
contribution to the process mining area, since it focusses on the usability of the
developed platform instead of describing the workflow monitoring process.

5.3 Process mining – enhancement type

( J-2) An intelligent production workflow mining system for continual quality
enhancement (2006).
This paper proposes a system to mine data in a production workflow by integrating
OLAP, fuzzy logic and ANN. Its goal is to discover the necessary knowledge to
support decision making used to improve the workflow in real time. OLAP is used to
store and enable dimensional analysis of the workflow monitored data and generate
mining historical data performed by ANN. In turn ANN is applied as a prediction
module, trained from the monitored historical data of the workflow. From the model
created, ANN makes a prediction about the workflow quality. If the model detects
that the quality is out of control, the system suggests refining the relevant parameters in
the workflow, through an improvement model implemented with fuzzy rules. With the
result of the implementation and system testing, the authors conclude that both
BPMJ computational intelligence techniques – ANN and fuzzy logic – can be used jointly to
21,6 improve the quality of the production workflows, minimizing the defects of their
resulting products.
(C-1) Development of a distributed process mining system for reactive ion etching
enhancement (2006).
In this paper, the authors use a very similar approach to the study ( J-2), where
1406 OLAP is used to organize data concerning the execution of a manufacturing process
and ANN is used to learn quality standards and then monitor in real time the execution
of process instances. The evaluation and the quality of prediction allow including
improvements in the actual manufacturing process. Differently from the study ( J-2),
particularly in this case, a set of neural networks is required to evaluate the different
properties in question. Thus, different regression models are generated, and a decision
tree helps in choosing which one to apply in each situation. The results of this study
demonstrate, through a case study conducted in six months, that the proposed
approach enables to continuously analyze the problems monitored to achieve the
quality improvement of the manufacturing process, optimizing it.
(C-2) A fuzzy paradigm approach for business process intelligence (2006).
This paper proposes the modeling of business processes based on fuzzy theory, also
with supported by ANN. The authors propose a system that includes a soft computing
module applied to the BPM context resulting in a system directed to the soft business
process intelligence. The soft computing module supports the decision making based
on fuzzy rules which were extracted from a set of synthetic data generated based on
rules and logic related to a process model taken from the context of bank business to
process credit application. The ANN was used for extracting fuzzy rules. Thus, in this
case process mining is used to automate the extraction of knowledge related to the
process so that it can later be improved with the help of fuzzy rules.
(C-3) Process optimization of candy production based on data mining (2011).
This paper presents a system to improve the candy production process,
consequently increase the production efficiency and improve product quality. In this
system, ANN is used to optimize the production process parameters. However, the
authors do not explore in detail how the neural model was developed and tested,
presenting only some information on the ANN architecture and some results obtained
from the model application. Although the authors do not clearly describe how the log
events are used, there are explicit references to the existence and use of a data
warehouse created by the data collected from the candy production instance process.
The data warehouse is configured as a data source for the optimization performed by
the ANN, which makes decisions about improvements in the production process
parameter, which must then be adjusted by human operators.
( J-4) Mining event logs to support workflow resource allocation (2012).
This study uses SVM to compare performance and classification results with other
techniques. The overall objective of this study was to propose and validate a system
that works as resource manager allocated to workflows. The manager system is
modeled as a classification task and based on an “a priori” algorithm used to find
frequent patterns in workflow event logs. The identified patterns are used to generate
association rules, respecting the existing restrictions in the allocation of resources in
the workflow activities. Based on the association rules created by the algorithm, the
system is able to classify resources for a new activity created in a workflow,
automating the resource management task. Assuming that the system indicates the
best resource for a given situation, it is understood that this approach helps to improve
the workflow. The approach was experimented, through data mining tools in WEKA, ANN and
using the following classification algorithms: SVM, C4.5, ID3 and Naïve Bayes. support vector
The author’s approach, using association rules, achieved the best results with SVM and
C4.5 (with centesimal differences) in terms of classification accuracy. However,
machines
a disadvantage reported for SVM regards the mining time, that is, the time for
optimizing the SVM decision model.
(C-4) Data mining for “big archives” analysis: a case study (2013). 1407
This work addresses archival analysis process (records reading, counting, sorting,
assessing, searching and grouping) in order to improve it. To achieve this goal, the
authors combine association rule mining and SVM in an analysis procedure applied in
a case study to address the formal records of communication (called State Cables)
between the US Department of State and its embassies and consulates worldwide. First,
the study was aimed to classify the cables in order to arrange them in relation to
security classes. Such classification was time-dependent, as discovered by the authors
during the tests carried out on the records collection. The second phase of the study
was intended to describe the cables and how the relationship between the cables’
content and their security classification changed over time. The SVM was applied in
order to infer the classifier for the first phase of the study. Different cable
representation models (i.e. different combination of record descriptors) were tested and
required a slide window representation of the time, since the classification behavior
was dependent on the period when the cables were created. The best accuracy reached
by the SVM in this initiative was 92.39 percent, using a model representation for cables
composed by information of subject and traffic analysis by geography and subject[4]
and slide window representation. The second phase of the study was implemented
using association rules discovery algorithms.

6. Discussion
Based on the information presented in the previous sections we conducted some
analyses, whose the results may be useful for the community of researchers interested
in using ANN and SVM to perform process mining.
First, regarding the number of identified primary studies, there is still a limited
number of studies being conducted with the goal of applying ANN and SVM in process
mining. Considering that approximately 620 works were initially identified in the area
of process mining, the 11 final studies selected represent only about 2 percent of all this
entire area. With respect to a possible trend regarding the distribution during the
11-year period evaluated, as shown in Figure 2, it appears that two-thirds of the studies
were published in the last four years, after a four-year period without any publication,
which may indicate an apparent research retake and interest in this area.
Moreover, in the studies evaluated the authors usually report optimistic impressions
regarding the use of these techniques in the process mining context, including
reasonable numerical results, by supporting the production of useful knowledge for
the business areas involved in the studies. However, considering the point of view
of the computational intelligence area, an analysis of the test and metrics strategies to
evaluate the quantitative results presented in these studies reveals a degree of
immaturity of these works in the standardization of the evaluation methods and the
availability of benchmark data sets that allow the reproducibility of the studies.
In order to support the analyses and discussion further presented here, Table XI
presents a summary of the data previously presented in Tables VI-X. It allows
to visualize in an integrated way all the 11 primary studies regarding all the evaluated
 21,6
BPMJ

1408

Table XI.

mining type)
Classification

together (ANN
summarized of
primary studies
regarding all the

data mining task

type, and process
evaluated aspects

architectures, SVM

SVM learning type,

variations, ANN and
(Process mining
type) Discovery Conformance Enhancement
Clustering Categorical prediction Categorical prediction Numeric prediction – Categorical prediction
(Data mining type) analysis (or classification) (or classification) “regression type” (or classification)
(ANN/SVM Unsupervised
learning type) learning Supervised learning Supervised learning
(Technique type) ANN ANN SVM ANN SVM
(ANN architect./ Competitive Feedforward SVM for Quality
SVM variation) layer multilayer classification Feedforward multilayer SVM for classification score

J-1 X 8.0
J-2 X 8.5
C-1 X 5.5
C-2 X 4.5
J-3 X 5.5
C-3 X 4.0
J-4 X 9.5
J-5 X 10.0
C-4 X 10.0
J-6 X 8.5
J-7 X 9.5
Note: X means that the evaluated attributes (at the first line) were found in the evaluated primary studies (at the first column)
aspects together as defined through the three research questions, i.e.: process mining ANN and
type, data mining task type, and details of the technique used. For sake of simplicity, support vector
Table XI presents only the attributes, and the combinations of attributes, that
were used in the previous tables. For example, the ANN architecture “feedforward
machines
single-layer” is not present in Table XI, since there is no study presented in Table VI
classified in such an attribute. Also aiming to summarize all the results together,
Figure 3 presents a cross-analysis considering the type of process mining versus the 1409
type of data mining task and some information regarding the technique used in each
primary study.
The works identified in this SLR share a pattern in relation to the items that we
investigated through the specific research questions. Most of the papers are studies
that perform some form of process mining enhancement, mainly using ANN. The
following presents specific analyses for each of the research questions investigated in
this SRL, as well as some crossed SRL analysis.
Regarding the type of process mining treated in the primary studies assessed,
according to Table XI and Figure 3, approximately 55 percent of the papers refer to
enhancement, followed by approximately 27 percent that refer to conformance, and about
18 percent refer to discovery. Although this classification takes into consideration the
definitions proposed by van der Aalst (2011), most of the primary studies identified in
this SLR did not fully adhere to the concepts presented by this process mining expert. For
example, only four of the 11 primary studies explicitly refer to the term process mining
( J-4, J-5, C-1, C-2) and one of them explicitly refers to the term workflow mining ( J-2).
Thus, slightly more than half of the identified primary studies were evaluated for
referring to the application of data mining in the business process context (or workflow),
which was understood as process mining indirectly. Additionally, only five of the 11
studies explicitly mention the use of logs in their approach, called event logs or process
logs, or instance logs, and other terms ( J-2, J-3, J-4, J-5, C-1). This means that slightly more
than half of the identified primary studies are not based on the analysis of event logs to
perform data mining tasks and may not accurately represent what is defined as process
mining. Finally, only five studies explicitly deal with process models or workflow models,
which is considered the basis for the three types of process mining according to van der
Aalst (2011) ( J-4, J-5, J-6, C-2, C-3). Overall, only two studies meet these three evaluated
characteristics together ( J-4 and J-5) (i.e. 22 percent) while three studies do not meet any
such characteristics ( J-1, J-7, C-4) (i.e. 33 percent).

Clustering analysis -
2
Unsupervised learning - Competitive layer ANN

Numeric prediction “Regression” type - 3

Supervised Learning - Feedforward multilayer ANN
Figure 3.
Categorical prediction (or classification) - 2
Summarized cross-
2
Supervised learning - SVM for classification analysis of the
primary studies
Categorical prediction (or classification) -
1 1 considering process
Supervised learning - Feedforward multilayer ANN mining type vs data
mining task type
and technique used
Discovery Conformance Enhancement
BPMJ In relation to the types of ANN and SVM used to implement data mining tasks when
21,6 specifically applied to the process mining context, the data presented show that:
• While 64 percent of the studies use ANN, only 36 percent of them use SVM. ANN
is more popular than SVM in various application areas, given that SVM began to
be studied more intensively only in the mid-1990s. ANN, however, has been
studied since the 1940s, although for most of the 1950s it has received very little
1410 attention from the academic environment. Given that SVM became popular so
recently, it is quite normal that there is a gap in its use compared to ANN in the
application areas, including process mining, although the two techniques show
similar results.
• As for the studies using ANN, 71 percent use the feedforward multilayer
architecture with supervised learning and only 29 percent use the competitive
layer architecture with unsupervised learning, and all with ANN SOM. The use of
these types of ANN is expected as they are the most popular in data mining,
mainly due to their potential problem solving, thereby a greater availability of
computational tools to support the application. However, the feedforward single-
layer, recurrent single-layer and recurrent multilayer architecture types, and also
the reinforcement learning and fixed-weight network types are not used in any of
the studies. This was also expected, since even in the data mining area in general,
these types of ANN appear less frequently, either because they are too simple and
limited – as the single-layer architecture, or because they are only suitable for
treating different types of problems – such as the reinforcement learning, highly
applicable in simulators and robotics.
• As for the studies using SVM, they all use the basic classification SVM model
(with supervised learning); that is, the variations SVR and SVC are not used in
any study. This is probably because ANN is more popular, since SVC and ANN
are similarly applicable for clustering contexts, like SVR and ANN for regression
contexts. At any rate, considering a total of only 11 primary studies, the lack of
studies that use SVC and SVR does not necessarily represent a pattern for the
process mining area, but rather only an indicative.
With regards to the data mining tasks being addressed in these process mining studies
which use ANN or SVM, the data presented show that: 55 percent of the identified
studies address the categorical prediction task (or classification); 27 percent address the
numeric prediction task (considering the “regression” type); and 18 percent address the
clustering analysis task. The numeric prediction task regarding the “trend
identification” type is not referenced in any of the studies. Regarding the frequent
patterns discovery, association and correlation rules discovery tasks, two studies
comment on the resolution of such tasks within the scope of a more comprehensive
approach, combined with categorical prediction (or classification) in one of the studies,
and with clustering analysis in another study. However, in those cases, this task is not
directly linked to the activities that ANN or SVM resolve, therefore do not fall within
the scope of this SLR analysis.
The cross-analysis presented in Table XI and Figure 3 showed that of the six studies
dealing with the categorical prediction task (or classification), four use SVM, while only
two use ANN. As for the numeric prediction tasks (considering the “regression” type)
all studies use ANN. Given that SVM is becoming increasingly popular and,
traditionally, knowing the SVC and SVR variations occur only later, it is likely that the
studies in the process mining area also start using SVC and SVR, instead of ANN, for ANN and
these two tasks which until now have been treated only with ANN. In general, SVM has support vector
often proved to be superior to ANN, considering the treatment of the structural risk,
besides the empirical risk which generates potentially more accurate results than
machines
treating only the empirical risk, as performed by ANN.
In general, approximately 80 percent of the primary studies evaluated in this SLR
showed positive results, with process mining benefits such as increased productivity 1411
and efficiency in business processes. However, not all these positive results can be
attributed to the use of ANN and SVN, since most studies suggest hybrid approaches
that combine a series of different techniques and technologies.
According to the evaluation of primary studies regarding their quality, as shown in
Table V, these studies were generally well evaluated, especially regarding the criteria
“reporting” and “relevance”. However, only papers published in the last three years
have been completely well evaluated regarding most of the criteria, including also
“rigor” and “credibility”, except for one study published in 2006, which in nearly all of
the criteria were also highly rated. The less recent papers refer to more basic studies,
and which do not clearly present the application, testing and verification of the results
achieved with the proposed approaches. These concentrate most of the studies
published in conference proceedings, which could somewhat explain the unfavorable
evaluation in terms of quality. Furthermore, the older papers refer to the period when
process mining was still being disseminated, with research work still focussed on
showing its potential. The good evaluation on the quality of the newly published works
shows that researchers are more concerned with the actual contribution of ANN and
SVM in the context of process mining.

7. Validity threats
In order to avoid threats to construct validity (Feldt and Magazinius, 2010), the “process
mining” and “workflow mining” terms were well established and are therefore sufficiently
stable to be used as search strings. Similarly, “business process” and “workflow”, together
with “data mining” terms, were also well established. Another point is to ensure
discovering all the primary studies in the theme chosen. For this purpose, two reputable
databases were used, namely, Scopus and ISI WoS. Together they index most of the
papers also indexed by other reputable databases, such as IEEEXplore, ACM Digital
Library, SpringerLink, ScienceDirect and Engineering Village.
As a threat to the internal validity (Alves et al., 2010), some subjective decisions may
have occurred during paper selection and data extraction since some primary studies
did not provide a clear description or proper purposes and results, making difficult the
objective application of the inclusion/exclusion criteria or the impartial data extraction.
To minimize selection and extraction systematic mistakes, several meetings were held
by the authors of this review work to discuss the type of proper treatment to the
identified conflicts. In order to minimize internal validity threats regarding data
analysis, this SLR uses some descriptive statistics techniques, such as: graphs to
summarize data; tabular description; and, parametric description (average).
Regarding external validity (Perry et al. 2000; Alves et al., 2010), we consider that
both scientific and industrial communities in the BPM and data mining domains can
benefit from the results of this SLR, and possibly the Information Systems and
computational intelligence fields as a whole. On the other hand, as a threat to external
validity, the scope of this SLR might not generalize to broader study selections other
than peer-reviewed papers, basically conference and journal papers.
BPMJ In terms of reliability purpose (Alves et al., 2010), we defined the search string and
21,6 procedures so that they could be directly and objectively replicated by other researchers.
Classification, however, is a source of reliability threat since evaluators inherently consider
subjective factors during primary studies analysis, even if carried out by several evaluators
in order to minimize distortions. Thus, there is no guarantee that other researchers could
achieve the same result as the primary studies classification presented herein.
1412
8. Conclusion
Through this SLR, it is found that although there is scientific interest in the area of
process mining, little has been specifically invested in ANN and SVM. In addition to the
relatively low number of primary studies identified, they do not cover the various
possible combinations in this context, not exploring the different existing possibilities
regarding the types of process mining or the types of architecture and learning related
ANN or to SVM variations. Considering the careful protocol we used for this SLR, is not
likely there are other studies that we did not identify here.
This scenario identified for the process mining context does not reflect the general
context of data mining, since, in this second case, ANN and SVM are widely used to solve
specific types of problems, as well as other computational intelligence techniques (Wang
and Fu, 2005; Wu et al., 2007). This low use of ANN and SVM may be due to a lack of
knowledge about their potential capability for this type of problem. Given that the area of
process mining involves multi-disciplinary knowledge, this may occasionally result in the
lack of expertise in certain areas of knowledge.
This SLR focussed only on ANN and SVM, which represent only about 2 percent of
all process mining studies currently developed and published. Other reviews must be
carried out with other computational intelligence techniques such as fuzzy set theory,
probabilistic reasoning evolutionary computing – including genetic algorithms and
genetic programing, etc. Our research group intends to investigate whether these other
techniques, also widely used in data mining in general, and which have already
presented good results, have been applied in process mining.

Notes
1. www.scopus.com
2. www.webofknowledge.com
3. Online Analytical Processing (OLAP): technology that uses a multidimensional view of
aggregate data to provide quick access to strategic information for further analysis. OLAP
enables analysts, managers, and executives to gain insight into data through fast, consistent,
interactive access to a wide variety of possible views of information (source: OLAP Council
White Paper - www.olapcouncil.org/research/whtpaply.htm).
4. TAGS: an identification system consisting of acronyms and words specifically designed to
organize communications disseminated throughout US Departments (source: www.state.gov/
documents/organization/89320.pdf).

References
Abonyí, J., Feil, B. and Abraham, A. (2005), “Computational intelligence in data mining”,
Informatica, Vol. 29 No. 1, pp. 3-12.
Alves, V., Niu, N., Alves, C. and Valenca, G. (2010), “Requirements engineering for software
product lines: a systematic literature review”, Information and Software Technology
(INFSOF), Vol. 52 No. 8, pp. 806-820.
Biolchini, J., Mian, P.G., Natali, A.C.C. and Travassos, G.H. (2005), “Systematic review in software ANN and
engineering”, Technical Report RT-ES 679/05, Federal University of Rio de Janeiro, Rio de
Janeiro, May.
support vector
Chen, L. and Ali Babar, M. (2011), “A systematic review of evaluation of variability management
machines
approaches in software product lines”, Information and Software Technology (INFSOF),
Vol. 53 No. 4, pp. 344-362.
Cortez, P. (2010), “Data mining with neural networks and support vector machines using the 1413
R/rminer tool*”, in Perner, P. (Ed.), Proceedings of 10th International Conference on Data
Mining, ICDM 2010, Springer-Verlag, Berlin, pp. 572-583.
Cortez, P., Cerdeira, A., Almeida, F., Matos, T. and Reis, J. (2009), “Modeling wine preferences by data
mining from physicochemical properties”, Decision Support Systems, Vol. 47 No. 4, pp. 547-533.
Cristianini, N. and Shawe-Taylor, J. (2000), An Introduction to Support Vector Machines and Other
Kernel-based Learning Methods, Cambridge University Press, Cambridge.
Dyba, T. and Dingsøyr, T. (2008), “Strength of evidence in systematic reviews in software
engineering”, in Rombach, H.D., Elbaum, S.G. and Münch, J. (Eds), Proceedings of the 2nd
International Symposium on Empirical Software Engineering and Measurement, ESEM
2008, ACM Press, Kaiserslautern, pp. 178-187.
Fausett, L. (1994), Fundamentals of Neural Networks Architectures, Algorithms and Applications,
Prentice Hall, Upper Saddle River, NJ, 461pp.
Fayyad, U., Piatetsky-Shapiro, G. and Smyth, P. (1996), “From data mining to knowledge
discovery in databases”, AI Magazine, Vol. 17 No. 3, pp 37-54.
Feldt, R. and Magazinius, A. (2010), “Validity threats in empirical software engineering research –
an initial survey”, Proceedings of the 22nd International Conference on Software
Engineering and Knowledge Engineering, SEKE 2010, Knowledge Systems Institute
Graduate School, Redwood City, CA, pp. 374-379.
Han, J. and Kamber, M. (2006), Data Mining – Concepts and Techniques, 2nd ed., Morgan
Kaufmann, Waltham, MA, 743pp.
Haykin, S. (2005), Neural Networks a Comprehensive Foundation, 2nd ed., Prentice Hall,
Upper Saddle River, NJ, 302pp.
Kitchenham, B. (2007), “Guidelines for performing systematic literature reviews”, Software
Engineering, Technical Report EBSE-2007-01, Software Engineering Group, School of
Computer Science and Mathematics, Keele University, Keele and Department of Computer
Science, University of Durham, Durham, July 9.
Nazem, S.M. and Shin, B. (2002), “Data mining: new arsenal for strategic decision-making”, in Becker,
S. (Ed.), Data Warehousing and Web Engineering, IRM Press, Hershey, PA, pp. 103-112.
Perry, D.E., Porter, A.A. and Votta, L.G. (2000), “Empirical studies of software engineering:
a roadmap”, Proceedings of the 22nd International Conference on Software Engineering,
Future of Software Engineering Track, ICSE 2000, ACM Press, Limerick, June 4-11,
pp. 345-355.
Stahl, F. and Jordanov, I. (2012), “An overview of the use of neural networks for data mining tasks”,
Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, Vol. 2 No. 3, pp. 193-208.
Tiwari, A. and Turner, C.J. (2008), “A review of business process mining: state-of-the-art and
future trends”, Business Process Management Journal, Vol. 14 No. 1, pp. 5-22.
van der Aalst, V.M.P. (2011), Process Mining – Discovery, Conformance and Enhancement of
Business Processes, Springer-Verlag, Heidelberg, Berlin, 352pp.
van der Aalst, W.M.P., Hofstede, A.H.M.T. and Weske, M. (2003), “Business process management:
a survey”, Proceedings of the International Conference on Business Process Management –
BPM, pp. 1-12.
BPMJ Vapnik, V.N. (1995), The Nature of Statistical Learning Theory, Springer-Verlag, Berlin and
Heidelberg, 314pp.
21,6
Wang, L. and Fu, X. (2005), Data Mining with Computational Intelligence, Springer-Verlag, Berlin
and Heidelberg, 276pp.
Weske, M. (2012), Business Process Management: Concepts, Languages, Architectures, 2nd ed.,
Springer-Verlag, Berlin and Heidelberg, 403pp.
1414 Wu, X., Kumar, V., Quinlan, F.R., Ghosh, J., Yang, Q., Motoda, H., McLachlan, G.J., Ng, A., Kiu, B.,
Yu, P.S., Zhou, Z.-H., Steinbach, M., Hand, D.J. and Steinberg, D. (2007), “Top 10 algorithms
in data mining”, Knowledge Information Systems, Vol. 14 No. 1, pp. 1-37.

Further reading
Esteva, M., Felix, J.T., Xu, W. and Anantha, K.P. (2013), “Data mining for ‘big archives’ analysis: a
case study”, in Bouthillier, F. and Ju, B. (Eds), Proceedings of the Association for Information
Science and Technology, ASIS&T 2013, ASIS&T Press, Montreal, pp. 90.01-90.10.
Ho, G.T.S., Lau, H.C.W., Ip, A.W.H. and Ning, A. (2004), “An intelligent information infrastructure
to support the streamlining of integrated logistics workflow”, Expert Systems, Vol. 21 No. 3,
pp. 123-137.
Ho, G.T.S., Lau, H.C.W., Lee, C.K.M., Ip, A.W.H. and Pun, K.F. (2006), “An intelligent production
workflow mining system for continual quality enhancement”, International Journal of
Advanced Manufacturing Technology, Vol. 28 No. 7, pp. 792-809.
Kang, B., Lee, J., Kim, D. and Kang, S.-H. (2011), “KNNI-SVM: a hybrid algorithm integrating
imputation and support vector machine for real-time business process monitoring”, ICIC
Express Letters – An International Journal of Research and Surveys, Vol. 5 No. 8(B)
pp. 2863-2868.
Kranjc, J., Smailovića, J., Podpečana, V., Grčara, M., Žnidaršiča, M. and Lavrača, N. (2014), “Active
learning for sentiment analysis on data streams: methodology and workflow
implementation in the ClowdFlows platform”, Information Processing and Management,
Vol. 51 No. 2, pp. 187-203.
Liu, T., Cheng, Y. and Ni, Z. (2012), “Mining event logs to support workflow resource allocation”,
Knowledge-Based Systems, Vol. 35, November, pp. 320-331.
Seret, A., Broucke, S., Baesens, B. and Vanthienen, J. (2014), “A dynamic understanding of
customer behavior processes based on clustering and sequence mining”, Expert Systems
with Applications, Vol. 41 No. 10, pp. 4648-4657.
Song, M., Yang, H., Siadat, S.H. and Pechenizhiy, M. (2013), “A comparative study of
dimensionality reduction techniques to enhance trace clustering performances”, Expert
Systems with Applications, Vol. 40 No. 9, pp. 3722-3737.
Thomas, O., Adam, O., Leyking, K. and Loos, P. (2006), “A fuzzy paradigm approach for business
process intelligence”, Proceedings of the 8th IEEE International Conference on E-Commerce
Technology (CEC 2006) and the 3rd IEEE International Conference on Enterprise
Computing, E-Commerce and E-Services (EEE 2006), IEEE Computer Society, Palo Alto,
CA, June 26-29, pp. 27.
Tsang, K.F., Lau, H.C. and Kwok, S.K. (2006), “Development of a distributed process mining
system for reactive ion etching enhancement”, Proceedings of the IEEE International
Conference on Industrial Informatics (INDIN 2006), IEEE Press, Singapore, August 16-18,
pp. 282-287.
Zhang, J.L., Yang, J.G., Shen, S.G. and Chen, H.Y. (2011), “Process optimization of candy production
based on data mining”, Advanced Materials Research, Vols-282-283, July, pp. 662-665.
About the authors ANN and
Ana Rocío Cárdenas Maita is a Graduate Student of the Master of Science Program in support vector
Information Systems at the University of São Paulo, Brazil. Bachelor in Computer Engineering
and Systems (2011) at the National University of San Antonio Abad del Cusco, Peru. Her main machines
research interests are information technology and business process management.
Lucas Corrêa Martins is an Undergraduate Student of the Bachelor Program in Information
Systems at the University of São Paulo, Brazil.
Dr Carlos Ramón López Paz is an Assistant Professor at the Higher Polytechnic Institute “José 1415
Antonio Echeverría”, Cuba. PhD in Technical Sciences (2011) at the University of Alicante, Spain;
Master in Applied Computing (2006) and a Bachelor in Computer Engineering (2006) at the
Higher Polytechnic Institute “José Antonio Echeverría”, Cuba. Dr. Carlos Paz has worked as a
Visiting Researcher (2013) at the University of São Paulo, Brazil. His main research interests are
consulting in business processes management, design science in information systems, strategic
alignment between business and IT, and IT governance.
Dr Sarajane Marques Peres is an Assistant Professor at the University of Säo Paulo, Brazil. PhD
in Electric Engineering (2006) at the University of Campinas; Master of Manufacturing Engineering
(1999) at the Federal University of Santa Catarina; a Bachelor in Computer Science (1996) at the
State University of Maringá, Brazil. She worked as Assistant Professor at the State University of
Western Paraná (1998-2005) and at the State University of Maringá (2005-2007), Brazil. Currently,
She is the tutor of the PET Information Systems USP group, under the Tutorial Education Program
of the Ministry of Education, Brazil. Her main research interests are computational intelligence,
data mining, machine learning, pattern recognition and gesture analysis.
Dr Marcelo Fantinato is an Associate Professor at the University of Säo Paulo, Brazil. PhD in
Computer Science (2007) and Master of Engineering (2002) at the University of Campinas;
Bachelor in Computer Science (1999) at the State University of Maringá, Brazil. He worked as a
Software Testing Specialist at the CPqD Foundation (2001-2006) and as a specialist in research
and development at Motorola (2006-2008), Brazil. He has a Green Belt certification in the Motorola
Six Sigma program for process improvement. His main research interests are business process
management, service-oriented computing, software product line, and software testing.
Dr Marcelo Fantinato is the corresponding author and can be contacted at: [email protected]

For instructions on how to order reprints of this article, please visit our website:
www.emeraldgrouppublishing.com/licensing/reprints.htm
Or contact us for further details: [email protected]