Proceedings of the 55th Hawaii International Conference on System Sciences | 2022

Improving Support Ticket Systems Using Machine Learning:

A Literature Review

Simon Fuchs Clemens Drieschner Holger Wittges

Technical University of Munich Technical University of Munich Technical University of Munich
[email protected] [email protected] [email protected]

Abstract and expensive [3]. On the other hand, it mostly

increases the ticket resolution time and therefore
Processing customer support requests via a support lowers the satisfaction of the customer initially
ticket system is a key-element for companies to provide creating the ticket [1, 5, 6].
support to their customers in an organized and At the same time, the volume of support tickets in
professional way. However, distributing and IT-companies created by customers has significantly
processing such tickets is much work, increasing the grown due to the digitalization efforts currently made
cost for the support providing company and stretching across all industries [7]. This means that IT-companies
the resolution time. The advancing potential of face an increasing pressure in automating their STSs
Machine Learning has led to the goal of automating to cope with the rising volume of tickets [7], to
those support ticket systems. Against this background, increase customer satisfaction [1, 8], to accelerate
we conducted a Literature Review aiming at support management processes [8, 9], and to reduce
determining the present state-of-the-art technology in costs [4, 10].
the field of automated support ticket systems. We With Artificial Intelligence (AI) and Machine
provide an overview about present trends and topics Learning (ML) algorithms becoming commonplace,
discussed in this field. During the Literature Review, the automating of STSs has become more interesting
we found creating an automated incident management than ever before [11]. Technologies for automated
tool being the majority topic in the field followed by ticket classification and automated ticket resolution
request escalation and customer sentiment prediction using ML open the possibility of automating basic
and identified Random Forrest and Support Vector day-to-day IT tasks replacing the first level support
Machine as best performing algorithms for staff members [4, 8].
classification in the field. This background in mind, we analyze current
trends and topics in automation of STSs using ML by
undertaking a narrative Literature Review [12]. We
1. Introduction
follow the principles of Watson and Webster [13] to
identify the relevant literature and to analyze the
Providing technical support for own IT products
present state of the art of automating STSs in the latest
is an integral part of software developing or software
scientific literature. In the process, we defined the
providing companies [1, 2]. For this purpose, most
following research questions to guide our Literature
companies provide their customers support ticket
Review: What is the present state-of-the-art
systems (STSs), in which users can create incident
technology in automating STSs? Which ML algorithms
tickets describing their problem or request [1]. In most
have the highest accuracy in classifying support
state-of-the-art STSs, at least some key decisions in
tickets?
distributing these support tickets to the responsible
In answering these questions, we aim to provide
support assistant or support team are still made by
an overview of the technical status quo of ML-driven
support staff members [3]. Support ticket distribution
automation of STSs. Further, we aim to identify
has the potential to bind a lot of working time of
research gaps in the current state of the art.
technically skilled workers, wherefore big companies
During our literature search, we recognized that
often use less-skilled or temporary workers for support
only few Literature Reviews in this field of research
ticket distribution or outsource the support to a third
have been published, not providing a general overview
party entirely [4]. This process of manually
of the field. With this Literature Review we want to
distributing emerging support tickets by often less-
provide such a general overview over the present
skilled human workers is on the one hand ineffective

URI: https://hdl.handle.net/10125/79570
978-0-9981331-5-7 Page 1893
(CC BY-NC-ND 4.0)
technological state of the art and the current status quo 3. Literature Review Design
of this particular field of research. Further, we want to
provide a paper for newcomers in the field to start Our main objective was to investigate the status
with. quo in the field of automating STSs using ML, to
understand state-of-the-art technology and to identify
research gaps in this field of research.
2. Background The Literature Review presented in this paper was
performed during February 2021. We searched the
STSs mostly enable customers to create a support databases Scopus, IEEE Xplore, Ebsco and Web of
incident. The term “support incident” denotes the Science. We limited the search to papers in English
whole entity of one single support process. Such language. It fast became obvious that precise search
support incidents mostly comprise a support ticket, words had to be found to confine the field of ML-
including a title, the plain text of the ticket and special automated STSs from more general fields like support
information, often called meta-data, like the priority of system improvement without ML or theoretical papers
the ticket, the category of the incident, a ticket-id, etc. concerning topics like Natural Language Processing
Furthermore, such incidents can comprise files or Deep Learning. In the end, we used some keywords
attached to a ticket or further customer data collected connecting them in various ways using the Boolean
by the system [1, 6, 8]. If we speak of “incident Operators AND and OR. The search words we used
management tools” in the course of this paper, we refer were “service desk”, “support ticket”, “support
to a concrete IT artifact that is able to manage support tickets”, “Machine Learning”, “ML”, “Artificial
incidents within a STS. In this context, managing an Intelligence”, “AI” and “classification”.
incident means: answering the incident, asking Depending on each database, we found another
questions to the creator of it or distributing it to a combination of those keywords to be useful to find
responsible support agent. relevant literature. For example, in Scopus we needed
The term “Machine Learning” has always more keyword to confine the search, whilst in the Web
denominated a very broad field of technical solutions of Science database fewer keywords lead to relevant
aiming at making “intelligent” machines [14]. In the results. The keywords finally used for each database
context of this Literature Review, the term “Machine are shown in Table 1. Unfortunately, it was not
Learning” is used for technologies comprising possible to find one general search string working in
algorithms, mathematical models and approaches that all databases due to the limits and specifications of
enable an IT artifact to automatically classify support these databases.
incidents based on previously provided training data. In total, we initially received 122 hits, from which
At this point, we want to highlight the differences we regarded 60 hits as relevant. Those 60 hits regarded
between chatbots and STSs. Chatbots are thought for as relevant comprised several duplicates. Eliminating
24-hour, real-time customer support, whereat mostly duplicates, we received 41 relevant papers. Doing
very simple questions are meant to be answered [15]. Forward and Backward Search [13] we additionally
STSs are meant to be a communication tool between found 2 papers we regarded as relevant. Google
customers and technical agents to solve technical Scholar was consulted during both.
problems occurring at customer side [1, 5]. Mostly, a The criteria for a relevant hit comprised: The main
support agent has to actively do something to solve the topic of the paper was a STS (the keyword “ticket”
customer request, for example install a program, sometimes produced hits in the field of tickets for air
modify a firewall, unlock ports, etc. or the question travel or festivals); Machine Learning, or at least
asked by the customer is difficult and one or more Data Science was actually used to improve a STS;
support agents have to think about or have to Technical solutions were developed, discussed, tested
investigate in order to solve the request [1]. Simply or deployed.
said, a chatbot is meant to answer simple customer
questions or for customer guidance [16], a STS is 4. Findings of the Literature Review
meant to help customers to create more difficult,
technical requests [8]. For this reason, we think that 4.1. Paper Type and Publisher
beside the growing hype around chatbots the field of
STSs will stay an independent and relevant field of A first finding of our Literature Review was that
research. a broad spectrum of publishers published the papers
found. Most papers (26 papers) found in the literature
search were published on conferences.

Page 1894
Table 1: Overview of search keywords and Table 3: Concept matrix
search results
IM Requ. Sent.
Paper Other
Rele- Tool Escal. Pred.
Databases Search Keywords Hits vant Asres, Mengistu [17] x x
Papers Nayak, Rai [15] x x
( "support ticket" OR Yang [18] x
"support tickets" OR Amin, Lancaster [19] (x)
"service desk") AND Baresi, Quattrocchi
x x
Scopus ("Machine Learning" OR 53 26 [20]
"AI" OR "ML") AND ( Han and Sun [21] x
"classification" ) Mour, Dey [22] x x
Revina, Buza [23] x
("All Metadata":support Xu, Mu [24] x x
ticket) AND Al-Hawari and
("All x
Barham [8]
Metadata":classification) 17 13 İşcen and Gürbüz [2] x
AND Lo, Tiba [25] x x
("All Metadata":Machine Mukunthan and
IEEE Learning) x
Selvakumar [26]
Explore ("All Metadata":service Nayebi, Dicke [6] x x
desk) AND Mandal, Agarwal [11] x
("All Misra and Podder [27] (x) x
Metadata":classification) 5 5 Palacios, Morillas
AND (x) x
[28]
("All Metadata":Machine
Shanmugalingam,
Learning) x
Chandrasekara [29]
TI = (ticket and
1 1 Werner, Li [30] x x x
classification)
Gajananan, Loyola
Ebsco TI = (incident management (x) x
[31]
AND Machine Learning OR 2 2
Gupta, Asadullah [1] x x
artificial intelligence)
Han, Goh [32] x
Web of TI=(support ticket*) 14 6
Koehler, Fux [33] x x
Science TI=(service desk *) 30 8
Lyubinets, Boiko [34] (x) x
In Total 122 60
Meng, Xu [35]) x
Without Duplicates 39
Montgomery, Damian
Articles identified in Forward & Backward (x) x (x)
2 [5]
Search
Paramesh, Ramya
Final Selection 41 x
[36]
Parmar, Biju [37] x
Patidar, Agarwal [38] x
Saberi, Theobald [39] x
Table 2: Overview of paper type and
Silva, Pereira [9] x
publishers Stein, Flath [7] x
Qamili, Shabani [3] x x x
Paper type Number of Hits in Percent Zhou, Zhu [40] x
Journal 17 39,5% Zuev, Kalistratov [10] x
Conference 26 60,5% Xu, Zhang [41] x
Publisher Number of Hits in Percent Chagnon [4] x
IEEE Conference 11 26,8% Giurgiu, Wiesmann
x
IEEE Journal 5 12,3% [42]
ACM 2 4,8% Montgomery and
x
Other Journal 11 26,8% Damian [43]
Other Conference 12 29,3% Reddy, Reddy T [44] x
Goby, Brandt [45] x x
(x) connotes a mentioning as minor topic or only
implicitly. Categories in the columns: development/
deployment/evaluation of an incident management
tool; customer request escalation prediction;
sentiment prediction, and Other. Papers are sorted
according to their publishing date from recent to prior.

Page 1895
 Over 35% of all found papers (16 papers) were between 80-90% for SVM seems reasonable [1, 2, 8,
published on IEEE conferences or in IEEE Journals. 9, 18, 37].
As written later below, most papers pertain a certain, The second place (7 papers) is taken by (versions
practical use case. We therefore conclude that the topic of) the Random Forrest (RF) approach that performed
of STSs automation is mainly interesting for the best in 3 papers. Here, we found maximal accuracies
application-oriented community that prefers to publish of 78% [30], 90% [6] and 92 % [36].
at conferences. All findings pertaining paper type and As it is shown in Table 4, RF performed especially
publisher are shown in Table 2. well in request escalation prediction and sentiment
prediction, while in ticket classification SVM
4.2. Topic Analysis performed better.
Table 4: ML algorithms and ML solutions
In a next step, we analyzed and categorized the evaluated and compared in the literature
topics dealt with according to Corbin and Strauss [46].
Especially, we were interested in technical features ML Algorithm Ticket Requ. Sent.
that are developed, tested, and deployed in the present Class. Escal Pred.
research. These findings are presented in Table 3. K-Nearest Neighbor 4 (0)
The topics discussed the most throughout the classification (KNN)
relevant literature are: development, deployment and Support Vector 9 (6) 2 (0) 1 (0)
evaluation of an incident management tool, customer Machine (SVM)
request escalation prediction, and customer sentiment Decision Tree (DT) 4 (0)
prediction. Apart from that, the topics examined in the Rule-Based 3 (0)
literature were quite individual. Papers that examined Naïve Bayes 6 (0) 2 (0) 1 (0)
further individual topics were (also) categorized as Random Forrest (RF) 6 (2) 2 (2) 1 (1)
Other. DNN 6 (3)
unsupervised learning 1 (0)
4.3. Machine Learning algorithms used <60%
accuracy
One of the questions that we were the most
The number in each cell indicates how often an
interested in was Which Machine Learning algorithms algorithm was evaluated for a specific task. The
are already in use and which have already proven number in brackets behind indicates how often the
themselves to work? 17 of the examined 41 papers algorithm was graded the best performing in the
implemented and evaluated at least one ML algorithm paper. The abbreviations for the tasks indicate (from
or solution known from literature that is more left to right): ticket classification, customer request
theoretical. 15 of these compared at least two of such escalation prediction, customer sentiment prediction.
algorithms/solutions with each other.
To determine the “best performing” algorithm in Additionally, Naïve Bayes approaches were
those papers, we first looked if the authors of each presented and evaluated in the literature (7 papers) but
paper identified one of their algorithms as “best no paper was found in which a Naïve Bayes algorithm
performing”. If this was not the case, we compared the could beat an SVM or an RF algorithm.
algorithms presented by their reported accuracy, In addition, a trend can be identified over the
precision and recall. In case of similarly performing years. According to Revina, Buza [23], the earliest
results, we weighted the reported accuracy as approaches of support ticket classification were ruled-
tiebreaker. A tabular overview of our findings based approaches that were outperformed by RF and
regarding used ML algorithms and how often they SVM. This probably is the reason that mostly in the
were the best performing ones in a paper is presented years 2018 and 2019, RF and SVM were the
in Table 4. approaches most implemented and evaluated and also
The ML algorithm most used (10 papers) for best performing [2, 6, 8, 9, 30, 36]. Since 2019, some
support ticket classification was (sometimes a authors also propose artificial-neuronal-network-
modified version of) the Support Vector Machine based solutions, because of the good performance of
(SVM) approach, which was also the algorithm that neuronal network approaches in other fields of ML
best performed in most papers (6 papers). In these 6 [19, 21, 23, 29]. While there is some promising
papers accuracies between 63% [37] and 98% [18] evidence, that Deep Neuronal Network (DNN) based
were reached. Those results are heavily depending on solutions might be able to outperform present ML
the data set used, but overall the papers an accuracy approaches like RF or SVM [29] (reported accuracies:
77% [29], 83% [19], 99% [21]), this point seems not

Page 1896
to be fully reached yet [23]. Revina, Buza [23] find in customers are not satisfied and escalate their request
their conclusion (page 11) “that simple algorithms by pressuring the support agent or the agents’
work well if using appropriate linguistic features” and supervisors. An escalation of a customer request
can be equally performative. mostly occurs when the customer creating the request
When analyzing the 4 papers presenting DNN- or a support manager is dissatisfied with the way a
based approaches [19, 21, 23, 29], we found that two support ticket is processed and requests its escalation.
of those papers used specific LSTM-networks, which This request escalation leads to a concentration of
were both the best performing approaches in that extensive human resources for solving the customer
papers [19, 29]. Since two papers are not a sufficiently request [30] within the support providing company.
large sample, no general statement can be derived This concentration of human resources means more
here. Nevertheless, LSTM DNNs seem to be a stress for the involved support agents and support
promising approach to which further attention should managers and additionally, these escalations are
be paid. mostly expensive for the company and significantly
lower customer satisfaction [30, 43].
4.4. Incident Management Tool Five papers of the analyzed literature deal with
ML-supported or even ML-driven escalation process
To develop, implement, test and deploy an optimization. The primary goal is often to predict the
automated tool for incident management is by far the risk of a support request to escalate, such that a
topic most treated in the found literature. 31 of the in company can concentrate resources on a customer
total 41 examined papers presented a complete, or at request before the customer gets angry and escalates
least components for a practical support ticket his/her request [5, 43]. Montgomery, Damian [5]
classification artifact. The remaining 10 papers developed a prototype that is able to predict escalation
discussed or evaluated only parts of such tools or probability in percent using Random Forest and
treated support ticket classification in a more XBoost algorithms reaching an accuracy of 81% on
theoretical approach. We take this finding as evidence their deployed use case. The artifact ESMMArT
that the majority of research in the field is focused on presented in Nayebi, Dicke [6] is also able to predict
creating practical technical applications. an escalation probability using Random Forrest
In most cases, a practical use case was on hand, reaching 90% accuracy. Further, Werner, Li [30]
such that the authors mainly developed such an artifact developed a cost-based mechanism to train and
for or optimized an existing one in a distinct situation evaluate ML algorithms for request escalation
(for example in [1]; [8]; [6]; [29]; or [17]). This of prediction.
course leads to a problem of generalizing their results
as already ascertained by Misra and Podder [27]. The 4.6. Sentiment Prediction
accuracy data presented by those papers should
therefore also be read with caution. The field of customer request escalation is closely
Misra and Podder [27] also stated that theoretical connected to the field of customer sentiment
ML knowledge obtained by the scientific literature prediction. As mentioned above, customer satisfaction
often has problems in unfolding its full potential when is a very important good when running an IT related
being applied in practical technical solutions, because business. As also mentioned, the quality of support
in reality either training data is not accessible like in services heavily influences customer satisfaction [1,
experiments or real-world data is more noisy or 30]. Therefore, it would be very useful for companies
diverse than polished data for university experiments. to be able to predict the sentiments of their customers,
Across all literature analyzed, there is a while a support incident is open [30]. In this context,
consensus, that training data is the linchpin of the term sentiment describes the general feelings of a
developing a well-functioning ML tool. In such, it is customer and his overall attitude towards the company
not surprising that every practical use case of and its support services. Mostly in the literature, the
developing, deploying, testing and optimizing a real customer’s sentiment is classified in the three
support ticket is massively impacted by the training categories “positive”, “neutral” and “negative” [30].
data accessible to the authors and the ticket data that Some of the analyzed papers deal with prototypes
are meant to be classified. for customer sentiment prediction. The main goal is
hereby to predict the sentiment that a customer is
4.5. Request Escalation Prediction experiencing during a support process, mostly after
he/she created a support ticket [3].
The term “Request Escalation” denotes the The predicted sentiment of a customer can then be
phenomenon that during a support process sometimes used for a wide scope of application: Gajananan,

Page 1897
Loyola [31] developed an artifact predicting customer The findings and results in these topics were as
sentiment from support ticket data for predicting the diverse as the topics itself. Nevertheless, we were able
probability if a customer would renew his subscription to carve out some general findings:
at a cloud business service; Werner, Li [30] use
customer sentiment prediction to predict the risk of  As in nearly any research in the field of
support escalation; and Qamili, Shabani [3] want to Machine Learning the accessibility and quality
use customer sentiment prediction for internal of training data importantly influences the
evaluation of the support system. outcome of the project [1, 3, 25]
The results in the field of emotion/sentiment  The metrics precision and recall are by far the
prediction are mixed. Qamili, Shabani [3] report low most-used metrics for evaluating ML ticket
accuracy (<45%) due to the difficulties in classification tools [5, 6, 18, 36].
accumulating labeled data for sentiment prediction.  Classification tools work more precisely the
Gajananan, Loyola [31] report high accuracy for their fewer classes they have to classify to [1, 3, 25].
subscription renewal prediction, but do not directly
predict customer sentiment. Instead, they only use 4.8. Data sets used
sentiment polarity as a parameter of their model.
Werner, Li [30] report pleasing successes using the Most papers analyzed in this Literature Review
Watson NLU model, but give no values for their used own datasets consisting of ticket data acquired in
emotion prediction. Overall, the impression arises that their specific use case. Often, papers were written in
so far only the surface of ML-based customer cooperation with companies providing ticket data for
sentiment prediction has been scratched. training and automated STSs were developed for these
companies. Because of non-disclosure agreements,
4.7. Specific topics according to the special these data sets are not available for the community and
use case - the category other also not described in detail in the papers analyzed.
Within the small group of remaining papers, we
Every paper analyzed during this Literature could not identify any publically available data set
Review is heavily influenced by the practical problem used in more than one paper.
it was written around. The presented artifacts and
prototypes were always developed and adjusted for a 5. Discussion
special use case and therefore, many papers at hand
face, present and solve very special problems. As a The primary goal of this Literature Review was to
result, in addition to the dominant topic of automated provide an overview over the field and to identify the
incident management tools, a large number of other present state of the art of automating STSs using
topics are dealt with or broached. A selection of these Machine Learning. Our first finding was that several
reads: prototypes of ML-automated STSs were developed,
 Automate labeling of tickets, either for training deployed and evaluated in the past 5 years. Indeed, to
or for easier ticket resolution by a support agent develop, deploy and often evaluate a practical
[34] prototype for a specific use case is the dominating
 Chatbots [15, 33, 44] topic within the field of ML-automated STSs. At this
 Spam detection [3] juncture, every prototype presented is heavily shaped
 Performance optimization [24] by the specific use case it was developed for. This
 Automated analysis of pictures attached to a leads to an overall problem of generalizability and
support ticket [11] validity of the results presented in each paper [1, 6].
 Business/process/text mining for better support Therefore, on the one hand, more comparative
system architecture [25, 45] research and meta-analysis of prototypes developed is
 AI explainability in support ticket automating needed. On the other hand, we understand the specific
[22] constraints every practical use case brings along.
 Ticket resolution time prediction [10] Technical limitations like the volume of accessible
 Automated STSs in context of Internet of training data or the volume of topics discussed within
Things (IoT) [28] a STS; special requirements like the number of
 Using answering bot (Microsoft LUIS) for categories to classify tickets in, the number of support
automated request responses [29] teams working at a company, a maximal response time
requested by the management or technical standards
within a company; and individual features like a

Page 1898
request escalation pipeline, a spam detector, a explainability (226 hits in IEEE database), chatbots
company-specific evaluation metric or the like all have (606 hits), IoT (57k hits) and performance
to be engaged individually. Therefore, we expect to optimization (132k hits) are highly investigated.
see more prototypes for specific use cases presented in Therefore, it is not surprising that these topics were
future research work. also researched in the context of support ticket
Regarding the variability of algorithms applied automating.
and tested in the papers, it appears clear that the Closing, we want to point out four topics that we
specific use case each paper deals with heavily considered being research gaps after analyzing the
influences which algorithms performs the best in each found literature:
paper. Therefore, in order to find the “best First, we found that in most papers, automating an
performing” algorithm in one specific use case always IT support desk is regarded as a process that reduces
fine-tuning and dealing with the specific use case will labor cost and increases customer satisfaction [1-4].
be necessary. Nevertheless, a statement can be made However, the perspective of support agents and
about which ML algorithms and solutions have proven support managers and how ML-optimized support
themselves during the past 5 years not in one paper desks can improve their working life is missed in the
only, but in various papers and hence in various use literature. Often, problems like angry customers, false
cases. Generally, ML algorithms like SVM and RF ticket allocation, frequently asked questions, etc. stress
have proven to be more accurate and precise than or annoy support agents and support managers the
“older” rules-based approaches or Decision Trees [2, most. Unfortunately, there is no research investigating
8, 23]. Also, there is some evidence that DNNs could how ML-driven automating can increase the job
outperform the present best-performing algorithms satisfaction of support agents and support managers,
SVM and RF, especially in the case of large training especially in the case of smaller companies with only
data and many classes to classify tickets in [21, 29]. limited human resources and the challenge of using
SVM and RF can therefore be considered standard their skilled employees as 2nd level support agents and
approach algorithms for support ticket classification. productive workforce at the same time.
Nevertheless, so far there is only pioneering work Second, in most papers analyzed the ML models
in the field of Deep Learning approaches for support were trained by human-guided training with data
ticket classification. Only 4 of the 41 analyzed paper manually labeled, e.g. in [1], [2], [6], or [8]. Manually
treated Deep Learning in the context of support ticket labeling thousands of tickets is mostly an unthankful
classification. More research is needed to evaluate if work, therefore it would be nice to apply unsupervised
Neuronal Networks can outperform present ML learning solutions for training data labeling. Actually,
algorithms in support ticket classification. Although, Lo, Tiba [25] did some pioneer work in this field.
there is a lot of research ongoing in the field of Deep Unfortunately, their unsupervised Kmeans and
Learning, the application of those results in STSs DBSCAN algorithms did not work very well. Also, the
should be intensified. Closing, we want to argue that unsupervised algorithms tested by Nayebi, Dicke [6]
there is also a need for more prototype-based research did not perform very well. Revina, Buza [23] used a
developing DNN-based, automated support ticket semi-supervised algorithm and comes to the finding:
desks. “semi-supervised learning (SSL) allows inducing a
Respective the topic of customer model from a large amount of unlabeled data
sentiment/emotion prediction, we see a great potential combined with a small set of labeled data” [23, page
for further research work. Knowing customer 4] that was promising but still needed some labeled
sentiment has big economical potential [30] either by data. We conclude that more research in this field is
helping customer requests not to escalate [5], by needed to lower the effort for labeling training data
helping to prioritize requests [6] or by helping to enormously.
predict if customers would renew their subscriptions Third, we found the approach presented by
[31]. But, multiple other use cases for knowing Shanmugalingam, Chandrasekara [29] very
customer sentiment can be thought. Nevertheless, interesting, where the authors of the paper created a
research in the field of predicting customer sentiment ticket classification tool, in which the determined class
based on their created support tickets is only few (see information are transmitted to a solution bot in order
above). Therefore, more research in this field is to generate a STS that automatically answers customer
needed. requests. As already said, there is a lot of research
Pertaining the topics categorized as Other, most going on pertaining chatbots and question-answer-
of the presented topics in the category Other above are systems are also a ML field that is researched in Jiang,
relevant and highly researched topics in the field of AI Su [47] and Shevchenko, Eremin [48]. Surprisingly,
and ML in general. Especially, topics like AI most papers analyzed during this Literature Review do

Page 1899
not use the class information received from their understanding, text generation or text processing. In
incident management tools to automate ticket addition, literature pertaining ML question answering
answering. Especially in the case of commonly asked systems were not considered. This of course means
questions or so-called first-level support requests, that research gaps identified above might be solved
automated ticket answering can save costs for the theoretically by research in those other fields.
company and resolution time for the customers. Nevertheless, the practical application of such
Therefore, we argue that more research in the field of theoretical known solutions to the field of STSs is a
automated ticket answering should be performed. promising task not realized yet.
Fourth, we recognized that all papers examined The Literature Search did only comprise
only focus on the positive impact that ML-automated searching scientific databases. We deliberately did not
STSs have on companies and customers. Meanwhile, search google, patent literature or other non-scientific
several studies have revealed skepticism of people publications. For this reason, we did not examine the
towards the usage of ML or AI (for example Gherheș present state of the art of commercial industry
and Obrad [49] or Morikawa [50]) and there are also solutions like ServiceNow or SAP Service Ticket
reports about negative effects of the usage of ML Intelligence, hence these were not examined in the
solutions in companies, like customer being around analyzed literature.
80% less satisfied and less likely to purchase when
they realize they communicate with a chatbot [16]. 7. Conclusion
Sometimes, there is a report of models not performing
that well yet (for example in Lo, Tiba [25]), but we Support ticket help desks are an essential part of
miss a treatment of negative aspects of ML-automated modern companies. As developments in the field of
STSs, like customers not trusting the systems, support ML advance, it becomes interesting to automate
agents not understanding the decisions of the system, support ticket help desks using ML solutions to lower
etc. Therefore, we argue for more attentiveness error rate and cost within the support department. In
regarding possible negative effects of ML-driven order to investigate the present state of the art in the
automation of STSs and especially we suggest field of ML-driven support ticket automating we
publishing such negative experiences and not only the conducted this Literature Review.
promising results. We found that every paper analyzed is heavily
Additionally, we argue that STSs are an essential influenced by the practical use case it was written
part of IT-related business and the potential of cost- about and that this is due to the individual nature of
reduction, raise in customer satisfaction and raise in support ticket systems, the companies they are used in
employee satisfaction through ML-driven automating and the special, technical requirements raised by ML
of these STSs makes the field not only interesting for algorithms. We also found that the majority of the
Computer Science Research or Engineering but papers in the field is published on conferences.
especially for Information Systems Research. We found the topic of creating an automated
Finally, in our personal research in the field, we incident management tool is the dominating topic in
found that customer guidance while ticket creation can this field of research. ML algorithms like RF and SVM
help improve the quality of trainings data, while also are currently best performing in ticket classification,
increasing customer satisfaction and increasing the while there is evidence that Deep Learning algorithms
accuracy of the ticket classifiers. Regarding the papers will take the lead in the near future. Two other
analyzed in this Literature Review, this topic seems important topics in the field found in this Literature
not investigated yet. Review were Request Escalation prediction, in which
For the future the authors of this paper aim at it is the goal to optimize processes within the support
doing more research in the topics of customer system handling customers escalating their requests,
guidance for better data, unguided machine learning and customer sentiment prediction, in which it is the
and (semi-)automated question answering. In goal to predict the sentiment of a customer based on
particular, the authors aim at creating an own STS for the tickets he/she created. Additionally, we found a
their own specific use case. great variety of topics individually dealt with in single
papers.
6. Limitations of this Literature Review We provide an overview over the current
technological state of the art, and give suggestions in
We intentionally narrowed the scope of this which directions the research scope can be expanded,
Literature Review to the topic of STSs automation. and identify research gaps.
This means that we did not consider literature in the
broader fields of ML-driven text classification, text

Page 1900
8. References [15] Nayak, S.P., A. Rai, K. Vankataramanappa, J.A.
Pansuriya and J. Singler. Adaptive MoD Chatbot:
Toward Providing Contextual Corporate Summarized
[1] Gupta, M., A. Asadullah, S. Padmanabhuni and A.
Document as Suggestions and Reported Issue Ticket
Serebrenik, Reducing user input requests to improve IT
Routing. in Advances in Artificial Intelligence and Data
support ticket resolution process. Empirical Software
Engineering. 2021. Singapore: Springer Singapore.
Engineering, 2018. 23(3): p. 1664-1703.
[16] Luo, X., S. Tong, Z. Fang and Z. Qu, Frontiers:
[2] İşcen, E.S. and M.Z. Gürbüz. A Comparison of Text
Machines vs. Humans: The Impact of Artificial
Classifiers on IT Incidents Using WEKA. in 2019 4th
Intelligence Chatbot Disclosure on Customer
International Conference on Computer Science and
Purchases. Marketing Science, 2019. 38(6): p. 937-947.
Engineering (UBMK). 2019.
[17] Asres, M.W., M.A. Mengistu, P. Castrogiovanni, L.
[3] Qamili, R., S. Shabani and J. Schneider, An Intelligent
Bottaccioli, E. Macii, E. Patti and A. Acquaviva,
Framework for Issue Ticketing System Based on
Supporting Telecommunication Alarm Management
Machine Learning. 2018. 79-86.
System With Trouble Ticket Prediction. IEEE
[4] Chagnon, C.T., A.C.; Djamasbi, S., Creating a decision
Transactions on Industrial Informatics, 2021. 17(2): p.
support system for service classification and
1459-1469.
assignment through optimization, in AMCIS 2017 -
[18] Yang, L., Fuzzy Output Support Vector Machine Based
America's Conference on Information Systems: A
Incident Ticket Classification. IEICE Transactions on
Tradition of Innovation. 2017, Worcester Polytechnic
Information and Systems, 2021. E104.D(1): p. 146-151.
Institute, Germany.
[19] Amin, K., G. Lancaster, S. Kapetanakis, K.-D. Althoff,
[5] Montgomery, L., D. Damian, T. Bulmer and S. Quader,
A. Dengel and M. Petridis. Advanced Similarity
Customer support ticket escalation prediction using
Measures Using Word Embeddings and Siamese
feature engineering. Requirements Engineering, 2018.
Networks in CBR. in Intelligent Systems and
23(3): p. 333-355.
Applications. 2020. Cham: Springer International
[6] Nayebi, M., L. Dicke, R. Ittyipe, C. Carlson and G. Ruhe,
Publishing.
ESSMArT way to manage customer requests. Empirical
[20] Baresi, L., G. Quattrocchi, D.A. Tamburri and W.-J.
Software Engineering, 2019. 24(6): p. 3755-3789.
Van Den Heuvel. Automated Quality Assessment of
[7] Stein, N., C. Flath and C. Böhm, Predictive Analytics for
Incident Tickets for Smart Service Continuity. in
Application Management Services. 2018.
Service-Oriented Computing. 2020. Cham: Springer
[8] Al-Hawari, F. and H. Barham, A machine learning based
International Publishing.
help desk system for IT service management. Journal of
[21] Han, J. and A. Sun. DeepRouting: A Deep Neural
King Saud University - Computer and Information
Network Approach for Ticket Routing in Expert
Sciences, 2019.
Network. in 2020 IEEE International Conference on
[9] Silva, S., R. Pereira and R. Ribeiro. Machine learning in
Services Computing (SCC). 2020.
incident categorization automation. in 2018 13th
[22] Mour, V., S. Dey, S. Jain and R. Lodhe. Feature Store
Iberian Conference on Information Systems and
for Enhanced Explainability in Support Ticket
Technologies (CISTI). 2018.
Classification. in Natural Language Processing and
[10] Zuev, D., A. Kalistratov and A. Zuev, Machine
Chinese Computing. 2020. Cham: Springer
Learning in IT Service Management. Procedia
International Publishing.
Computer Science, 2018. 145: p. 675-679.
[23] Revina, A., K. Buza and V.G. Meister, IT Ticket
[11] Mandal, A., S. Agarwal, N. Malhotra, G. Sridhara, A.
Classification: The Simpler, the Better. IEEE Access,
Ray and D. Swarup. Improving IT Support by
2020. 8: p. 193380-193395.
Enhancing Incident Management Process with Multi-
[24] Xu, J., J. Mu and G. Chen, A multi-view similarity
modal Analysis. in International Conference on
measure framework for trouble ticket mining. Data &
Service-Oriented Computing (ICSOC). 2019.
Knowledge Engineering, 2020. 127: p. 101800.
[12] Paré, G., M.-C. Trudel, M. Jaana and S. Kitsiou,
[25] Lo, D., K.K. Tiba, S. Buciumas and F. Ziller. An
Synthesizing information systems knowledge: A
Emperical Study on Application of Big Data Analytics
typology of literature reviews. Information &
to Automate Service Desk Business Process. in 2019
Management, 2015. 52(2): p. 183-199.
IEEE 43rd Annual Computer Software and
[13] Watson, R.T. and J. Webster, Analysing the past to
Applications Conference (COMPSAC). 2019.
prepare for the future: Writing a literature review a
[26] Mukunthan, M.A. and S. Selvakumar, Multilevel Petri
roadmap for release 2.0. Journal of Decision Systems,
net-based ticket assignment and IT management for
2020. 29(3): p. 129-147.
improved IT organization support. Concurrency and
[14] Stone, P.B., Rodney; Brynjolfsson, Erik; Calo, Ryan;
Computation: Practice and Experience, 2019. 31(14): p.
Etzioni, Oren; Hager, Greg; Hirschberg, Julia;
e5297.
Kalyanakrishnan, Shivaram; Kamar, Ece; Kraus, Sarit;
[27] Misra, J. and S. Podder. Mining Text in Incident
Leyton-Brown, Kevin; Parkes, David; Press, William;
Repositories: Experiences and Perspectives on
Saxenian, AnnaLee; Shah, Julie; Tambe, Milind; Teller,
Adopting Machine Learning Solutions in Practice. in
Astro;, "Artificial Intelligence and Life in 2030.".
2019 34th IEEE/ACM International Conference on
Report of the 2015-2016 Study Panel, Stanford
Automated Software Engineering Workshop (ASEW).
University, Stanford, CA, September 2016, 2016.
2019.

Page 1901
[28] Palacios, D., C. Morillas, M. Garcés and P. Tapia. Big customer recognition in contact centers: Dealing with
data-empowered system for automatic trouble ticket common names. Expert Systems with Applications,
generation in IoT networks. in 2019 IEEE 2nd 5G 2018. 113: p. 356-376.
World Forum (5GWF). 2019. [40] Zhou, W., X. Zhu, T. Li, L. Shwartz and G.Y.
[29] Shanmugalingam, K., N. Chandrasekara, C. Hindle, G. Grabarnik. Multi-view feature selection for labeling
Fernando and C. Gunawardhana. Corporate IT-Support noisy ticket data. in NOMS 2018 - 2018 IEEE/IFIP
Help-Desk Process Hybrid-Automation Solution with Network Operations and Management Symposium.
Machine Learning Approach. in 2019 Digital Image 2018.
Computing: Techniques and Applications (DICTA). [41] Xu, J., H. Zhang, W. Zhou, R. He and T. Li, Signature
2019. based trouble ticket classification. Future Generation
[30] Werner, C., Z.S. Li and D. Damian, Can a Machine Computer Systems, 2018. 78: p. 41-58.
Learn Through Customer Sentiment?: A Cost-Aware [42] Giurgiu, I., D. Wiesmann, J. Bogojeska, D. Lanyi, G.
Approach to Predict Support Ticket Escalations. IEEE Stark, R.B. Wallace, M.M. Pereira and A.A. Hidalgo,
Software, 2019. 36(5): p. 38-45. On the adoption and impact of predictive analytics for
[31] Gajananan, K., P. Loyola, Y. Katsuno, A. Munawar, S. server incident reduction. IBM Journal of Research and
Trent and F. Satoh. Modeling Sentiment Polarity in Development, 2017. 61(1): p. 9:98-9:109.
Support Ticket Data for Predicting Cloud Service [43] Montgomery, L. and D. Damian. What do Support
Subscription Renewal. in 2018 IEEE International Analysts Know About Their Customers? On the Study
Conference on Services Computing (SCC). 2018. and Prediction of Support Ticket Escalations in Large
[32] Han, R., V.K.H. Goh, A. Sun and M. Akbari, Towards Software Organizations. in 2017 IEEE 25th
Effective Extraction and Linking of Software Mentions International Requirements Engineering Conference
from User-Generated Support Tickets. 2018. 2263- (RE). 2017.
2271. [44] Reddy, Y., C. Reddy T and P. Dayaker, To dialogue
[33] Koehler, J., E. Fux, F.A. Herzog, D. Lötscher, K. with chatbot using machine learning. International
Waelti, R. Imoberdorf and D. Budke. Towards Journal of Mechanical Engineering and Technology,
Intelligent Process Support for Customer Service 2017. 8: p. 729-739.
Desks: Extracting Problem Descriptions from Noisy [45] Goby, N., T. Brandt, S. Feuerriegel and D. Neumann.
and Multi-lingual Texts. in Business Process Business Intelligence for Business Processes: the Case
Management Workshops. 2018. Cham: Springer of IT Incident Management. in ECIS. 2016.
International Publishing. [46] Corbin, J.M. and A. Strauss, Grounded theory research:
[34] Lyubinets, V., T. Boiko and D. Nicholas. Automated Procedures, canons, and evaluative criteria.
Labeling of Bugs and Tickets Using Attention-Based Qualitative Sociology, 1990. 13(1): p. 3-21.
Mechanisms in Recurrent Neural Networks. in 2018 [47] Jiang, C., Q. Su, L. Zhang and B. Huang, Automatic
IEEE Second International Conference on Data Stream Question Answering System Based on Convolutional
Mining & Processing (DSMP). 2018. Neural Network and Its Application to Waste Collection
[35] Meng, F.J., J. Xu, X. Zhang, L. Yang, P. Chen, Y. System. Journal of Circuits, Systems and Computers,
Wang, X. Liu, N. Ayachitula, K. Murthy, L. Shwartz, 2021. 30(01): p. 2150013.
G. Galambos, Z. Su and J. Zheng. Opportunities and [48] Shevchenko, A.P., R.A. Eremin and V.A. Blatov, The
Challenges Towards Cognitive IT Service Management CSD and knowledge databases: from answers to
in Real World. in 2018 IEEE Symposium on Service- questions. CrystEngComm, 2020. 22(43): p. 7298-
Oriented System Engineering (SOSE). 2018. 7307.
[36] Paramesh, S.P., C. Ramya and K.S. Shreedhara. [49] Gherheș, V. and C. Obrad, Technical and Humanities
Classifying the Unstructured IT Service Desk Tickets Students’ Perspectives on the Development and
Using Ensemble of Classifiers. in 2018 3rd Sustainability of Artificial Intelligence (AI).
International Conference on Computational Systems Sustainability, 2018. 10(9): p. 3066.
and Information Technology for Sustainable Solutions [50] Morikawa, M., Firms' Expectations about the Impact of
(CSITSS). 2018. AI and Robotics: Evidence from a Survey. Economic
[37] Parmar, P.S., P.K. Biju, M. Shankar and N. Kadiresan. Inquiry, 2017. 55(2): p. 1054-1063.
Multiclass Text Classification and Analytics for
Improving Customer Support Response through
different Classifiers. in 2018 International Conference
on Advances in Computing, Communications and
Informatics (ICACCI). 2018.
[38] Patidar, M., P. Agarwal, L. Vig and G. Shroff,
Automatic Conversational Helpdesk Solution using
Seq2Seq and Slot-filling Models, in Proceedings of the
27th ACM International Conference on Information
and Knowledge Management. 2018, Association for
Computing Machinery: Torino, Italy. p. 1967–1975.
[39] Saberi, M., M. Theobald, O.K. Hussain, E. Chang and
F.K. Hussain, Interactive feature selection for efficient

Page 1902