DOI: 10.3303/CET23107044 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Paper Received: 12 April 2023; Revised: 26 April 2023; Accepted: 15 June 2023 
Please cite this article as: Erdős F., Farhat R., 2023, Sustainability Approach of SAP Application Management Service Solutions in the Field of 
Warehouse Management, Chemical Engineering Transactions, 107, 259-264  DOI:10.3303/CET23107044 
  

 CHEMICAL ENGINEERING TRANSACTIONS  
 

VOL. 107, 2023 

A publication of 

 

The Italian Association 
of Chemical Engineering 
Online at www.cetjournal.it 

Guest Editors: Petar S. Varbanov, Bohong Wang, Petro Kapustenko 

Copyright © 2023, AIDIC Servizi S.r.l. 

ISBN 979-12-81206-07-6; ISSN 2283-9216 

Sustainability Approach of SAP Application Management 

Service Solutions in the Field of Warehouse Management 

Ferenc Erdős*, Rahma Farhat 

Department of Informatics, Széchenyi István University, Győr, Hungary, 

erdosf@sze.hu 

Organizations using SAP systems encounter the challenge of enhancing services while minimizing costs and 

response times. To achieve this, SAP offers a comprehensive range of sustainable cloud solutions in the field 

of Application Management Services (AMS), integrating machine learning, big data, and smart-machine 

technologies. This research endeavors to offer an alternative, sustainable approach, utilizing non-cloud-based 

machine learning to improve the efficiency of ticket classification in SAP AMS. This study presents an SAP 

Extended Warehouse Management (EWM) related case study in which various machine learning algorithms 

were applied to real-world SAP ticket data to automatically categorize incident tickets. As warehousing plays a 

major role in streamlining the supply chain, this approach aligns with the goals of sustainable supply chain 

process optimization. During the analysis, various classification algorithms were compared to achieve the best 

metrics. Our research did not just analyze the different algorithms for that specific business problem. The best 

model was integrated and deployed as well to accomplish a sustainable SAP AMS solution in the field of EWM. 

1. Introduction 

In today's digital era, information technology (IT) plays a vital role in the operations of many organizations. As 

businesses increasingly rely on IT systems, managing and maintaining complex infrastructures, such as SAP 

applications, becomes crucial. However, these systems' growing size, variety, and sophistication pose 

challenges in terms of cost, complexity, and sustainability (Bisong, 2019). 

Sustainability is increasingly crucial in the business world, and SAP SE has taken a proactive approach to 

corporate sustainability by aiming to have a positive economic, social, and environmental impact. To achieve 

this, SAP offers a comprehensive range of sustainable cloud solutions. For instance, SAP offers a solution called 

Service Ticket Intelligence on the SAP Cloud BTP (Business Technology Platform), which employs machine 

learning for service ticket classification. However, many companies are not yet ready to adopt these cloud-

based solutions. 

This research paper focuses on the sustainability approach of SAP Application Management Service (AMS) 

solutions. AMS involves the strategic management of business application portfolios, including deployment, 

maintenance, support, and monitoring (Li et al., 2013). It aims to ensure high service quality and availability 

through regular service operations described in service-level agreements (SLAs) (Odun-Ayo et al., 2017). 

The escalating technological advancements, such as cloud computing, artificial intelligence, blockchain, and 

augmented reality, introduce new complexities and maintenance issues to IT infrastructures. Organizations must 

align their IT strategies with business objectives while controlling costs and optimizing resources. However, 

managing SAP development and maintenance expenses requires specialized knowledge and expertise that 

may not be readily available within a company. 

The paper aims to explore and propose a sustainable way to SAP AMS, considering the need for efficient 

resource management, cost control, and the adoption of cutting-edge technologies. By integrating sustainability 

principles, organizations can improve service quality, reduce costs, and enhance overall efficiency. 

Additionally, the research addresses the challenges of ticket classification within AMS for SAP applications. The 

manual categorization of support tickets is time-consuming and prone to errors, affecting response times and 

customer satisfaction. The paper suggests the creation of an artificial intelligence-powered ticket classification 

259



engine to handle this. This engine would automate the ticket-handling process, significantly reducing manual 

labor and minimizing human errors while enhancing service quality and user satisfaction. 

The objective is to minimize manual intervention, reduce categorization errors, and improve the overall process 

efficiency of warehousing, which is directly connected to supply chain management. By automating ticket 

categorization and assigning steps, the human resources of the IT help desk can be allocated to more valuable 

functions, thereby reducing Extended Warehouse Management (EWM) process breakdowns. With the 

improvement of warehouse management processes, supply chains can be more streamlined and can make 

them more sustainable. 

2. Related Work 

The automation of service desk ticket classification has been a focus of research in recent years, aiming to 

improve incident management teams' productivity, reduce ticket routing time, and minimize categorization 

errors. Several studies have explored different machine learning (ML) techniques and their applications in 

automating incident ticket classification. 

This research path is based on text classification via Natural Language Processing (NLP), which plays a key 

role in the digitization of a wide range of modern industries. Uysal and Gunal (2014) highlighted that choosing 

the right combination of preprocessing functions instead of applying or avoiding them all significantly improves 

classification accuracy. Their research used frequently used preprocessing procedures, like eliminating stop 

words, lowercase conversion, stemming, and tokenization. Howard and Ruder (2018) emphasized the impact 

of fine-tuning and directionality on the behavior of a classifier, which can increase the model's performance by 

0.5-0.7 times. Pranckevičius and Marcinkevičius (2017) – for basics and Shah et al. (2020) – advanced cases, 

compared several ML models such as Naive Bayes, Logistic Regression, Random Forest, Decision Tree, 

Support Vector Machines (SVM), and K-Nearest Neighbors (KNN), and provided their pros and cons for text 

classification purposes. Their findings indicate that for this application, the Logistic Regression multi-class 

classification achieved the highest accuracy. 

Service ticket classification-related papers are focusing on incident content classification. Silva et al. (2018) 

developed a mechanism using support vector machines (SVM) and KNN algorithms to automatically classify 

incident tickets. SVM achieved an overall accuracy score of 89 %, while KNN achieved 82 % accuracy, 

showcasing the effectiveness of ML approaches in incident categorization. Paramesh and Shreedhara (2019) 

proposed an automated system for classifying service desk tickets, employing supervised ML techniques such 

as logistic regression, KNN, Multinomial Naive Bayes (MNB), and SVM. The study highlighted the superiority of 

SVM in achieving accurate classification results compared to other models. 

Khramov (2018) conducted research comparing ML models, including SVM, Naive Bayes, and Random Forest, 

using grid search techniques to optimize hyperparameters. The SVM model, implemented with a Boolean vector 

representation, achieved an accuracy of 84.8 %. Altintas and Tantug (2014) suggested a two-phase 

categorization approach for incident tickets, utilizing Term Frequency-Inverse Document Frequency (TF-IDF) 

for feature extraction. Four algorithms, including KNN, SVM, decision trees, and Naive Bayes, were tested on a 

dataset of 10,000 problem tickets. SVM exhibited the highest accuracy of 86 % across various ML algorithms 

and datasets. Al-Hawari and Hala (2021) developed a technical support system employing an ML algorithm for 

efficient ticket classification. Utilizing ticket comments and descriptions during training improved model 

accuracy. Different ML models, including J48 (tree), Decision Table (rule), Naive Bayes, and SMO (SVM), were 

evaluated, with SMO achieving 81.4 % accuracy. 

From a scientific perspective, these studies collectively indicate that machine learning techniques offer 

promising opportunities for automating ticket classification processes. However, it is important to consider the 

specific characteristics of the dataset, the preprocessing methods employed, and the choice of evaluation 

metrics when implementing these techniques. 

Based on the literature review, we could not find a specific approach to classifying incident tickets for SAP 

Warehouse Management incident ticketing data. Furthermore, the reviewed studies only analyze the different 

metrics of the models and lack the deployment process of the models. 

3. Methodology 

To achieve the research goal, we applied the Cross-Industry Standard Process for Data Mining (CRISP-DM) 

framework as the guiding methodology for this project. Each phase of the methodology is adapted and tailored 

to suit the specific requirements and challenges of automating ticket classification (Figure 1). 

260



 

Figure 1: Flowchart of the methodology 

It may be necessary to alternate between various phases because the phases' order is not strictly followed. The 

flowchart's arrows show the most significant and common dependencies between steps. 

3.1 Phase 1: Business Understanding 

The first task of the CRISP-DM methodology is the Business Understanding phase (Chapman et al., 2000). In 

this phase, a comprehensive understanding of the business problem and requirements for automating ticket 

classification in SAP AMS is developed. During this phase, extensive research was conducted in collaboration 

with a company that offers SAP support services and specializes in assisting clients with SAP EWM modules. 

Through in-depth investigations and consultations with the company's personnel, it was discovered that they 

were facing a significant challenge in efficiently assigning the right tickets to the appropriate solution group. This 

inefficiency directly led to customer dissatisfaction as well as financial issues through breakdowns in warehouse 

processing. The existing manual ticket classification process proved to be time-consuming and error-prone, 

resulting in delayed resolutions and dissatisfied customers. The misallocated tickets often ended up in the hands 

of less qualified support staff, leading to suboptimal solutions and, in some cases, even escalations. The 

company's management emphasized the need for a robust and automated ticket classification system to 

streamline their operations, enhance customer satisfaction, and reduce financial losses resulting from misrouted 

or delayed tickets. By identifying these critical points and gathering insights from the company's operations, the 

foundation was laid for devising a suitable machine-learning solution that addresses the specific challenges in 

the realm of ticket classification for SAP AMS. 

3.2 Phase 2: Data Understanding 

The Data Understanding phase involves exploring and analyzing the dataset of support tickets in SAP AMS. 

The dataset consists of 1630 SAP EWM-related incidents, each with a unique key, creation date, requester, 

responsible person, and a title and description of the reported issue. The data has been modified to comply with 

GDPR regulations for privacy and security. 

The incidents are managed through the Jira ticketing system used by the company. They are manually created 

and assigned to solution groups. The dataset was collected from the IT help desk professionals, providing 

insights into the specific support levels assigned to each complaint. During data preprocessing, irrelevant 

columns were removed, and the focus was given to the ticket description in English and the corresponding level 

of support. The distribution of tickets among these levels is as follows: Level 0 - 46.6 %, Level 1 - 22.9 %, and 

Level 2 - 30.7 %. 

3.3 Phase 3: Data Preparation  

After the necessary data cleaning tasks, concrete steps were involved in the pre-processing of the textual data. 

Regular expressions were used to strip the text of all URLs, cutting down on clutter and unnecessary details. 

Tokenization was used to divide the text into smaller chunks, and lemmatization was used to reduce each word 

to the smallest possible unit. The improved language comprehension allowed the machine learning models to 

perform better. Oversampling methods were used to correct the asymmetry in the data. The dataset was made 

even more by resampling the underrepresented groups until they had the same amount of samples as the 

dominant group. The TF-IDF method was used to quantify textual information. This method weights terms 

according to how often they appeared in a certain document and how uncommon they were overall (Qaiser and 

Ali, 2018). These procedures guaranteed a high-quality dataset and set the stage for model creation and testing 

to follow. 

261



3.4 Phase 4: Modeling 

During the modeling phase, the available data is utilized to develop a predictive model (Lamba and 

Madhusudhan, 2022). This phase involves selecting appropriate modeling approaches, experimenting with 

various algorithms, and comparing their outcomes using evaluation metrics and graphical representations (Raju 

et al., 2023). The objective is to identify the optimal model that can improve accuracy and effectively address 

the problem at hand. 

In this phase, suitable modeling techniques are chosen based on the specific problem. This special case is a 

multi-class classification problem where the dataset consists of three distinct classes (L0, L1, and L2), and we 

utilized supervised machine learning techniques. These techniques consider the target feature as the decision 

to be predicted and the other features as input variables. In this study, the following modeling approaches and 

variants have been selected for comparison: 

• Logistic Regression Model (L2 – Ridge regression) 

• Decision Tree (CART – Classification and Regression Trees) 

• Random Forest (Gini impurity) 

• Support Vector Machine (C-SVC – C-Support Vector Classification) 

• Neural Network (Multilayer Perceptron, one hidden layer, containing 100 hidden units) 

These modeling techniques are widely used by data scientists for multi-class classification tasks (Pranckevičius 

and Marcinkevičius, 2017), and they provide a diverse range of algorithms to explore (Ejiyi et al., 2022). 

3.5 Phase 5: Evaluation 

After implementing the models, a comparative analysis is conducted to evaluate their performance. During the 

performance measurement, K-Fold Cross-validation was used with 5 folds to determine the averages of the 

metrics. The applied evaluation metrics provided an overview of each model's performance in terms of accuracy, 

F1 score, precision, and recall. These metrics help assess the models' ability to classify the tickets accurately. 

3.6 Phase 6: Deployment 

The final step in the data science process is deployment, where the trained model is put into production to 

integrate it into the decision-making process. In this phase, the model is applied to achieve the specified 

business goals of the company. During the deployment phase, the ticket classification process was automated 

by integrating the trained model with Jira. The trained model was serialized using the pickle module and saved 

as a file. This model file was then uploaded to the deployment script. A connection to Jira was established using 

the Jira API, and authentication was achieved using an email and API token. 

To automate the ticket classification process, a script was developed, which retrieved new IT tickets from Jira 

using a JQL query and classified them using the trained model as in Figure 2. The script checked if a prediction 

comment already existed for a ticket and skipped it if present. For tickets without a prediction comment, the 

relevant ticket information was obtained, and the model was used to classify the ticket. Comments containing 

the predicted categorization were added to the ticket. 

 

Figure 2: Automated EWM ticket classification workflow 

The deployment architecture consisted of three tiers: the client tier, server tier, and data tier. The client tier 

presented the user interface and connected user requests to the server tier. The server tier, an HTTP web 

server, communicated with the Jira API and processed user requests by invoking the ML model. The data tier 

directly interacted with the Jira database. 

4. Results 

The comparative analysis pointed out the differences between the performance of the applied machine learning 

models. After the evaluation phase, the results are as follows. 

262



Table 1: Comparing evaluation metrics 

Model Name Accuracy F1-score Precision Recall 

Logistic Regression 0.75 0.73 0.8 0.73 

Decision Tree 0.80 0.83 0.83 0.84 

Random Forest 0.83 0.82 0.83 0.83 

SVM 0.79 0.79 0.80 0.77 

Neural Network 0.79 0.81 0.82 0.82 

The results indicate that the linear random forest classifier achieves the highest accuracy at over 83 %. This 

performance is comparable to the human operator-produced error rates of ticket classification. To assess the 

performance of the models further, the Receiver Operating Characteristic (ROC) curve was utilized. The ROC 

curve illustrates the trade-off between true positives and false positives in a binary classifier system (Gajjalavari 

and Rudravaram, 2022). Despite having a multiclass classification problem, each class was considered 

separately to evaluate the machine learning algorithms (Dingle et al., 2022). 

The Area Under the Curve (AUC) of the ROC curve was compared for each model (Figure 3), serving as a 

synthetic performance index. Among the tested classification algorithms, the Random Forest exhibited the 

highest AUC of 0.96, indicating its superior performance compared to the other algorithms. The logistic 

regression, support vector machine (SVM), neural network, and decision tree algorithms achieved AUCs around 

0.92, 0.95, 0.94, and 0.85, respectively. 

 

Figure 3: Comparing evaluation metrics 

Based on the comparative study using evaluation metrics and ROC curves, it was observed that Random Forest 

and SVM reached the highest values. The SVM algorithm achieved an AUC of 0.95, showcasing its strong 

performance in the classification task. Considering the better evaluation metrics and higher AUC value achieved 

by the Random Forest algorithm, it was selected as the final predictive model. 

However, it is essential to acknowledge that the performance of different machine learning algorithms can vary 

depending on the dataset and the specific field of application. The deployment phase successfully automated 

the EWM ticket classification process by integrating the trained Random Forest ML model with Jira. The script 

executed the classification of new tickets and added predicted categorizations as comments. 

By streamlining the ticket routing process, the research results can enhance the operational efficiency of EWM 

processes. This generates a supply chain-related advancement in reduced delays and improved response 

times. Such operational efficiency improvements contribute to overall sustainability efforts by minimizing energy 

consumption and optimizing resource utilization. 

5. Conclusions 

In conclusion, this research has successfully contributed to the sustainability approach of SAP EWM Application 

Management Service through the development of a support prediction model based on ML techniques. As a 

novelty, this specific comparative analysis of Logistic Regression, Decision Tree, Random Forest, SVM, and 

Neural Network highlighted the variations in how well the used machine learning models performed in this SAP 

EWM ticketing dataset. The applied prediction model improved resource allocation by automating the ticket 

263



classification and routing process. By efficiently assigning IT resources and support personnel, unnecessary 

workload and resource wastage are reduced, leading to improved resource efficiency and sustainability. 

Consequently, the utilization of machine learning techniques and the analysis of large volumes of data enable 

data-driven decision-making within SAP AMS. This approach supports sustainable practices by providing 

valuable insights for optimizing operations, resource allocation, and service delivery.  

In summary, this research demonstrated the potential of machine learning algorithms to automate and optimize 

SAP EWM ticket classification processes in AMS, aligning with sustainability goals without relying on cloud 

solutions. By implementing these sustainable and efficient approaches, IT service organizations can enhance 

their services, reduce costs, and improve overall customer satisfaction. In future research, it is recommended 

to conduct a comprehensive analysis to quantify the specific environmental, social, and financial benefits 

achieved through the implementation of the support prediction model. This analysis will provide a deeper 

understanding of the sustainability impact of supply chain improvements and further strengthen the sustainability 

approach of AMS. 

References 

Al-Hawari F., Hala B., 2021, A machine learning based help desk system for IT service management. Journal 

of King Saud University-Computer and Information Sciences, 33(6), 702–718. 

Altintas M., Tantug A.C., 2014, Machine learning based ticket classification in issue tracking systems. In 

Proceeding Int. Conf. Artif. Intell. Comput. Sci.(AICS 2014), 195–207. 

Bisong E., 2019, The multilayer perceptron (MLP). Building Machine Learning and Deep Learning Models on 

Google Cloud Platform: A Comprehensive Guide for Beginners. APress, Berlin, Germany, 401–405. 

Chapman P., Clinton J., Kerber R., Khabaza T., Reinartz T., Shearer C., Wirth R., 2000, CRISP-DM 1.0: Step-

by-step data mining guide. SPSS inc, 9(13), 1–73. 

Dingle K., Batlle P., Owhadi H., 2023, Multiclass classification utilising an estimated algorithmic probability prior. 

Physica D: Nonlinear Phenomena, 448, 133713. 

Ejiyi C.J., Qin Z., Salako A.A., Happy M.N., Nneji G.U., Ukwuoma C.C., Gen J., 2022, Comparative analysis of 

building insurance prediction using some machine learning algorithms. International Journal of Interactive 

Multimedia and Artificial Intelligence, 7(3), 75–85. 

Gajjalavari S., Rudravaram V. V., 2022, Multi-Class Classification using Mixtures of Univariate and Multivariate 

ROC Curves. Journal of Biostatistics and Epidemiology, 8(2), 208–233. 

Howard J., Ruder S., 2018, Universal language model fine-tuning for text classification. In: Proceedings of the 

56th Annual Meeting of the Association for Computational Linguistics, ACL 2018 Volume 1, 328–339. 

Khramov D., 2018, Robotic and machine learning: how to help support to process customer tickets more 

effectively. BSc Dissertation, Metropolia University of Applied Sciences, Helsinki, Finland. 

Lamba M., Madhusudhan M., 2022, Predictive Modeling. In: Text Mining for Information Professionals: An 

Uncharted Territory. Springer International Publishing, Cham, Switzerland, 213–242. 

Li Y., Li T.H., Liu R., Yang J., Lee J., 2013, Application management services analytics. In: Proceedings of 2013 

IEEE International Conference on Service Operations and Logistics, and Informatics, 28-30 July, Dongguan, 

China, 366–371, DOI: 10.1109/SOLI.2013.6611442. 

Odun-Ayo I., Ajayi O., Omoregbe N., 2017, Cloud service level agreements–issues and development. In: 2017 

International Conference on Next Generation Computing and Information Systems (ICNGCIS), 11-12 

December, Jammu, India, 1–6, DOI: 10.1109/ICNGCIS.2017.18. 

Paramesh S.P., Shreedhara K.S., 2019, Automated IT service desk systems using machine learning 

techniques. In: Data Analytics and Learning: Proceedings of DAL 2018. Springer Singapore, 331–346. 

Pranckevičius T., Marcinkevičius V., 2017, Comparison of naive bayes, random forest, decision tree, support 

vector machines, and logistic regression classifiers for text reviews classification. Baltic Journal of Modern 

Computing, 5(2), 221–232. 

Qaiser S., Ali R., 2018, Text mining: use of TF-IDF to examine the relevance of words to documents. 

International Journal of Computer Applications, 181(1), 25–29. 

Raju M., Shanmugam K.P., Shyu C.R., 2023, Application of Machine Learning Predictive Models for Early 

Detection of Glaucoma Using Real World Data. Applied Sciences, 13(4), 2445. 

Shah K., Patel H., Sanghvi D., Shah M., 2020, A Comparative Analysis of Logistic Regression, Random Forest 

and KNN Models for the Text Classification. Augmented Human Research, 5, 1–16. 

Silva S., Pereira R., Ribeiro R., 2018, Machine learning in incident categorization automation. In: 2018 13th 

Iberian Conference on Information Systems and Technologies (CISTI), 13-16 June 2018, Caceres, Spain, 

1–6, DOI: 10.23919/CISTI.2018.8399244. 

Uysal A.K., Gunal S., 2014, The impact of preprocessing on text classification. Information Processing & 

Management, 50(1), 104–112. 

264