2024 International Conference on Science, Engineering and Business for Driving Sustainable Development Goals (SEB4SDG) | 979-8-3503-5815-5/24/$31.

00 ©2024 IEEE | DOI: 10.1109/SEB4SDG60871.2024.10630378

Optimizing Network Management and Virtualization

Using Machine Learning Approach: Network Slice
Prediction
Lateef Adesola Akinyemi Olamide Peter Oshinuga Sunday Oladayo Oladejo Stephen Obono Ekwe
Departments of Computer Department of Electronic and School for Data Science and Department of Elect., Electronic
Science and Elect. Engineering, Computer Engineering Computational Thinking and Computer Engineering
CSET, Lagos State University Stellenbosch University Cape Peninsula University of
University of South Africa Lagos, Nigeria. Stellenbosch, South Africa. Technology
Johannesburg, South Africa [email protected] [email protected] Cape Town, South Africa.
[email protected] [email protected]

Mbuyu Sumbwanyanbe Ernest Mnkandla Oluwagbemiga Omotayo Shoewu

Department of Electrical Engineering, Department for Computer Science, CSET Department of Electronic and Computer
CSET, University of South Africa Engineering
University of South Africa Johannesburg, South Africa Lagos State University
Johannesburg, South Africa [email protected] Lagos, Nigeria.
[email protected] [email protected]

Abstract— This study demonstrates the utilization of Machine

Learning (ML) for network slice prediction, enabling the
optimization of resources for diverse network slices. Traditional
methods for network slice prediction often lack efficiency and
result in inaccuracies. By leveraging ML algorithms such as Naive
Bayes and Random Forest, an intelligent framework that
automates network slice prediction is developed. This framework
enhances network virtualization and management, facilitating
resource allocation. The ML algorithms take real-time network
conditions and usages, such as packet delay and smart city, as Fig. 1. Concept of Machine learning.
input and output for selecting the most suitable network slice. Data
analysis is further conducted to reveal the connections between the Network management refers to the activities involved in
input parameters and how these parameters influence the controlling and optimizing a computer network. It includes
selection of the accurate network slice. Network slicing plays a
tasks such as configuration, performance monitoring, and
crucial role as it enables the customization of services and
facilitates efficient scaling to meet the specific needs of different troubleshooting [3]. Network slicing supports network
applications and industries. The accuracy scores of the employed management by enabling efficient resource allocation. As
ML algorithms were generally perfect, except for the KNN and illustrated in Fig. 2, different agents are assigned to manage
SVM classifiers, which achieved an accuracy of 94.30% and specific network slices. This approach enables tailored
92.16%, respectively, for the prediction of network slices based on management of individual slices, enhancing the scalability and
incoming network connections and usages. flexibility of the computer network while reducing maintenance
costs.
Keywords— Network Slicing, Network Virtualization, Network
Management, Machine Learning, Data Analysis.

I. INTRODUCTION
Machine learning (ML) involves the creation and utilization of
models or algorithms that learn from data they are trained with
[1]. After training, the algorithms are employed to make
informed decisions, such as predictions and solving complex
problems, as illustrated in Fig. 1.

Fig. 2. Network management enhancing scalability and flexibility.

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
Authorized licensed use limited to: University of Illinois. Downloaded on October 15,2024 at 14:04:20 UTC from IEEE Xplore. Restrictions apply.
 The concept of network slicing involves the creation of
virtual networks, known as slices, within a physical network
infrastructure, enabling customized network architectures for
different applications or services [2]. Each network slice is
often customized to cater to specific service requirements, as
depicted in Fig. 2. Slice 1 is a specialized segment designed for
large-scale users and devices, such as database servers. Slice 2
is a segment tailored for medium-sized users and IoT devices.
Slice 3 is a distinct segment customized for small-scale users
and applications, as shown in Fig. 3. This virtualization
technique enables network operators to allocate resources and
provide diverse services efficiently.

Fig. 4. Aim of the study showing the input and output predictions of the
employed machine learning algorithms.

Fig. 3. Showing the concept of network slicing. The integration of ML algorithms with network slicing offers
several potential benefits. It enables the efficient classification
The accurate prediction and classification of network slices of network resources into their appropriate slices by
in a communication network is a significant challenge. considering factors such as network performance, connection
Traditional methods for network slice prediction often lack types, and usage. The machine learning system accesses the
efficiency and result in inaccuracies, which in turn hinder the incoming connections and places them in their suitable slices,
optimal utilization of network resources and diminish network as illustrated in Fig. 5. This ensures the automatic selection of
performance. To address this challenge, there is a need to the most suitable network slice based on the type of incoming
develop a reliable and automated approach that leverages connections, preventing network downtime and minimizing
machine learning techniques. Employing machine learning disruptions to network services.
algorithms makes it possible to proactively predict and classify
network slices based on incoming network connections, usages,
and their unique characteristics.
Network slicing enables network operators to make better use
of their network resources by creating numerous logical
networks that coexist within the same physical infrastructure
[3]. Each slice can be customized to meet certain capacity,
latency, or other network parameters. This adaptability enables
the network to be dynamically assigned and optimized in
response to the fluctuating demands of various situations, as
shown in Fig.4. This study aims to utilize machine learning
algorithms such as Random Forest, Naive Bayes, and others to
predict and classify network slices. These algorithms are
trained with network performance metrics and network usages,
such as packet loss rate, packet delay, and more, to predict the
corresponding output representing one of the network slices
(slice 1, slice 2, or slice 3), as illustrated in Fig. 4. The trained
ML algorithms Identify and ensure the selection of the most
suitable network slice based on real-time network conditions
and usages.

Fig. 5. The role of machine learning algorithms in resource allocation.

Authorized licensed use limited to: University of Illinois. Downloaded on October 15,2024 at 14:04:20 UTC from IEEE Xplore. Restrictions apply.
 II. RELATED WORKS algorithms through predictions and evaluation using four
The research conducted in [4] delves into the realm of evaluation metrics.
network slicing techniques tailored specifically for 5G
networks. Also, the study in [5] emphasized network slicing as
a critical enabler for the future of the internet. The authors in [6]
employed machine learning algorithms, including LSTM and
RF, to predict the throughput for individual network slices
within a 5G network infrastructure based on the allocated
resources. Following these predictions, the effectiveness of the
employed algorithms was evaluated using NS-3. The findings
indicated that RF outperformed LSTM. The study in [7]
extended its application of deep learning models to anticipate
whether a network provider would satisfy a new network slice
request, taking into account the conditions of the network
channel. The utilized model reduced the occurrence of false
positive results by 75%. In [8], a prediction algorithm for
network slicing was created utilizing a multi-dimensional
neural network to predict the network resources needed by a
device based on historical network demands and the optimal
slice selection. The algorithm exhibited an accuracy rate of
98.70%. The study in [9] employed machine learning
algorithms to forecast the precise slice category. Furthermore,
it juxtaposed the results of these algorithms using assessment
criteria such as accuracy, precision, sensitivity, and more. The
authors in [10] proposed a method called DeepCog that uses
deep learning to autonomously manage resources and optimize
Fig. 6. Methodology process.
performance in 5G networks with network slicing. The
approach dynamically allocates resources and adapts to B. Data Analysis
changing network conditions, enhancing efficiency and user
This is done to uncover the relationship between each network
experience. The works in [11-13] equally focused on network
performance metric and each network slice. This helps us
slicing concerning wireless communications using
understand how network conditions and usages influence the
metaheuristic algorithms such as genetic algorithms, particle
selection of the most suitable network slice.
swarm optimization, and so on. The authors in [14 -15] utilized
machine learning for predictions. Furthermore, varieties of ML C. Machine Learning Algorithms/ Classifiers
and metaheuristic algorithms have been employed in studies The following ML algorithms are used for the multi-class
[18-20] to address a range of wireless-inspired challenges in the prediction of network slices (Slice 1, 2, or 3):
recent past. 1) Support Vector Machine (SVM): This algorithm finds
the hyperplane to separate the different classes in the data. SVM
III. METHODS also supports multi-class prediction [17]. The One-vs-All
method was employed for the multi-class prediction task in this
A. Data Collection study. In this approach, three binary classifiers (Classifiers 1, 2,
The dataset used in this study comprises 31,584 rows with 16 and 3) were created. Each classifier distinguishes a single slice
instances [16]. Among these, 4 instances pertain to network from the combination of the remaining two slices. Classifier 1
performance metrics, while the remaining 12 instances learns to differentiate Slice 1 from the combination of Slice 2
represent different network use cases. Network performance and Slice 3, and so on, as illustrated in Fig. 7. The predicted
metrics are indicators that evaluate the effectiveness and slice is the one with the highest decision score among these
stability of the network. It includes attributes such as packet classifiers.
loss rate, packet delay, and more. On the other hand, the
network use cases represent various scenarios where the
network is applied, which include health care, smart cities,
homes, and more. The 16 instances are utilized to predict the
target variable, which is the slice type, consisting of three
distinct values: slice 1, slice 2, and slice 3. The methodology
process is presented in Fig. 6. The dataset was initially analyzed
and then further partitioned into training data (70%) and testing
data (30%). The training data was utilized to instruct the ML
algorithms, enabling them to discern relationships among all
the attributes within the dataset. Subsequently, the testing data Fig. 7. SVM approach for multi-class prediction.
was used to assess the performance of the trained ML

Authorized licensed use limited to: University of Illinois. Downloaded on October 15,2024 at 14:04:20 UTC from IEEE Xplore. Restrictions apply.
 2) Gaussian Naive Bayes (NB): This is a probabilistic IV. RESULTS AND DISCUSSION
classifier which is based on the Bayes theorem. It creates a A. Data Analysis
classification model by calculating the conditional probabilities
for each input attribute. It presumes that all characteristics are 1) Network Usages: This refers to scenarios where the
impartial and equally affect the target variable. network is applied. Fig. 9 provides a visualisation of the
3) Random Forest (RF): This is essentially a group of network use cases for each of the three slices. slice 1 exhibits
decision tree classifiers, where each tree can vote for the the highest number of usages, primarily driven by smartphone
prediction class that it thinks would be most accurate. In this usage. slice 2 demonstrates usages related to smart cities and
study, we optimized the number of trees used in the RF model IoT devices. on the other hand, slice 3 highlights significant use
for multi-class prediction of slices. Through experimentation cases in public safety, smart transportation, and healthcare, with
and cross-validation, 100 trees were employed in the RF, which 2,000 instances each. Based on this analysis, it can be
provided the best balance for effective predictive performance. concluded that:
4) K-nearest neighbors (KNN): The 'K' parameter in KNN  Slice 1 is specifically designed for small-scale usage.
defines the number of neighbouring point values utilized in the  Slice 2 caters to medium-scale usages.
classification task. This method selects an unlabeled data item  Slice 3 is tailored to accommodate large-scale usage.
by examining its K closest neighbors and allocating it to the
most prevalent category represented among them. KNN finds
the five nearest neighbours for each point represented by (*) in
Fig. 8, as implemented in this study.

Fig. 9. Showing network use cases across each slice.

2) Network Performance Metrics: These metrics are

measures used to evaluate the effectiveness, performance, and
overall operation of a network.
a) Packet delay: Packet delay indicates the time for a
packet to be delivered from the sender to the receiver. Fig. 10
Fig. 8. KNN approach for multi-class prediction. provides insights into the packet delay characteristics of
different slices:
D. Model Evaluation
 Slice 1 shows a higher density of delays between 50 to
This is done using evaluation metrics to validate the predictions
140ms.
made by each of the employed algorithms for the prediction of
 Slice 2 exhibits a wider distribution of packet delays,
network slice. This study employs four evaluation metrics:
indicating higher variability.
1) Accuracy: This refers to the capacity of the classification
 Slice 3 does not display a significant relationship with
algorithm to forecast the expected result.
(TP+TN) packet delay, suggesting consistent and stable
Accuracy = (1) performance.
(TP +FP+TN+FN)
2) Sensitivity: This signifies the number of correct positive
predictions made by the model.
TP
Sensitivity = (2)
(TP+FN)
3) Precision: This indicates the ratio of accurate positive
outcomes concerning all potential outcomes.
TP
Precision = (3)
(TP+FP)
4) F-measure: This is commonly referred to as the F1
score, which represents the harmonic mean of precision and
sensitivity.
(Precision x Sensitivity)
F − measure = (4)
(Precision+Sensitivity)

Fig. 10. Showing the connection between packet delay and slices.

Authorized licensed use limited to: University of Illinois. Downloaded on October 15,2024 at 14:04:20 UTC from IEEE Xplore. Restrictions apply.
 In summary, Slice 1 experiences concentrated delays, Slice
2 has more variability, and Slice 3 demonstrates consistent
performance with minimal delay variations.

b) 5G/LTE Category: Fig. 11 highlights that Slice 3 is

associated with the highest value of 5G/LTE network
performance, indicating its superior performance in certain
aspects. In contrast, Slice 1 has the lowest value of 5G/LTE
network performance but demonstrates a denser distribution,
implying a more concentrated performance range.

Fig. 13. Confusion matrix for KNN classifier.

The confusion matrix for the SVM classifier is depicted in Fig.

14. Impressively, the classifier achieved perfect predictions for
4329, 2189, and 2214 occurrences of slice 1, slice 2, and slice
3, respectively. However, there were instances where the
classifier made errors. Specifically, it incorrectly predicted 743
occurrences of slice 2 when the true value was slice 1.

Fig. 11. Showing distribution of 5G/LTE across all slices.

B. Confusion Matrices
The confusion matrix for RF and NB classifiers exhibited
equal outcomes is presented in Fig. 12, with an equal
distribution across all the cells. This indicates the two classifiers
achieved perfect prediction results. Specifically, when
considering a total number of 5072 occurrences, slice 1 resulted
in perfect predictions for all of them. Similarly, for slice 2, there
were 2189 outcomes with perfect predictions and 2214 for slice
3.

Fig. 14. Confusion matrix for SVM classifier.

TABLE I EVALUATION METRICS SCORES FOR ALL ALGORITHMS

Algorithms Accuracy Precision Sensitivity F-measure
RF 100.00 100.00 100.00 100.00
SVM 92.16 91.55 95.11 92.53
KNN 94.30 93.40 96.43 94.46
NB 100.00 100.00 100.00 100.00

The result obtained in Table I is visualized in Fig. 15. The

visualization shows that RF and NB classifiers exhibited
exceptional performance in predicting slices, as assessed by the
Fig. 12. Confusion matrix for RF and NB classifiers. evaluation metrics. These two classifiers achieved a perfect
score of 100% across all four evaluation metrics, highlighting
In Fig. 13, the confusion matrix for the KNN classifier. The their effectiveness in slice prediction. While SVM and KNN
classifier accurately predicted 4534 occurrences of slice 1. also demonstrated strong performance, they fell slightly short
However, it made 2 incorrect predictions of slice 1 when the compared to the top two classifiers. KNN outperformed SVM
actual value was slice 2. Furthermore, the classifier predicted with an accuracy score of 94.20%, whereas SVM achieved an
538 occurrences of slice 2 when the true value was slice 1. On accuracy score of 92.16%. These findings emphasize the
a positive note, the classifier achieved perfect predictions for superior performance of RF and NB classifiers in predicting
2187 occurrences of slice 2 and 2214 occurrences of slice 3. slices, making them the most reliable choices for accurate slice
classification in this study.

Authorized licensed use limited to: University of Illinois. Downloaded on October 15,2024 at 14:04:20 UTC from IEEE Xplore. Restrictions apply.
 [6] J. S. Camargo, E. Coronado, B. Gomez, D. Rincon, and S. Siddiqui,
“Design of AI-based Resource Forecasting Methods for Network
Slicing,” 2022 International Wireless Communications and Mobile
Computing (IWCMC), May 2022, doi:
10.1109/iwcmc55113.2022.9824551.
[7] M. Toscano, F. Grunwald, M. Richart, J. Baliosian, E. Grampin, and A.
Castro, “Machine Learning Aided Network Slicing,” 2019 21st
International Conference on Transparent Optical Networks (ICTON), Jul.
2019, doi: 10.1109/icton.2019.8840141.
[8] J. Zhao and G. Peng, “Industrial Internet Network Slice Prediction
Algorithm Based on Multidimensional and Deep Neural Networks,”
Proceedings of the 2022 5th International Conference on Artificial
Intelligence and Pattern Recognition, Sep. 2022, doi:
10.1145/3573942.3573989.
[9] N. A. Mohammedali, T. Kanakis, A. Al-Sherbaz, and M. O. Agyeman,
“Traffic Classification using Deep Learning Approach for End-to-End
Slice Management in 5G/B5G,” 2022 13th International Conference on
Information and Communication Technology Convergence (ICTC), Oct.
Fig. 15. Visualization of evaluation metrics scores for each classifier. 2022, doi: 10.1109/ictc55196.2022.9952446.
[10] D. Bega, M. Gramaglia, M. Fiore, A. Banchs and X. Costa-Perez,
"DeepCog: Cognitive Network Management in Sliced 5G Networks with
V. CONCLUSION Deep Learning," IEEE INFOCOM 2019 - IEEE Conference on Computer
Communications, Paris, France, 2019, pp. 280-288, doi:
The study has shown that the integration of machine 10.1109/INFOCOM.2019.8737488.
learning algorithms has proven to be a valuable approach in [11] S. O. Oladejo, "Efficient radio resource management for the fifth
accurately predicting and classifying network slices. This generation slice networks,” Ph.D. Thesis, University of Cape Town, 2021.
enables efficient resource allocation and enhances the [12] S. O. Oladejo, S. O. Ekwe, and L. A. Akinyemi, "Multi-tier multi-domain
utilization of network resources. Data analysis revealed that network slicing: A resource allocation perspective," in 2021 IEEE
AFRICON, pp. 1-6, IEEE, 2021.
Slice 1, 2, and 3 are designed to cater to the needs of small-
[13] S. O. Oladejo, S. O. Ekwe, and L. A. Akinyemi, "Multi-tier multi-tenant
scale, medium-scale, and large-scale users and applications network slicing: A multi-domain games approach," [Online]. Available:
respectively. By leveraging machine learning techniques, it was https://www.itu.int/pub/S-JNL-VOL2.ISSUE6-2021-A04, 2021.
demonstrated that network slicing can be automated, leading to [14] L. A. Akinyemi, O. P. Oshinuga, O. A. Eshilokun, S. O. Oladejo and S.
better network management, network virtualization, and O. Ekwe, "Development of Phylum Chordata Sound Recognition System
enhanced network performance. Our analysis of various using Machine Learning and Deep Learning Techniques: A Case Study
of Catfish," 2023 International Conference on Electrical, Computer and
performance metrics demonstrated the effectiveness of Energy Technologies (ICECET), Cape Town, South Africa, 2023, pp. 1-
machine learning algorithms, such as RF and NB classifiers, in 6, doi: 10.1109/ICECET58911.2023.10389248.
predicting the perfect network slice for incoming connections [15] L. A. Akinyemi, O. P. Oshinuga, S. O. Ekwe and S. O. Oladejo,
accurately, with an accuracy score of 100% each. "Enhancing Chronic Kidney Disease Prediction Through Data
Preprocessing Optimization and Machine Learning Techniques," 2023
International Conference on Electrical, Computer and Energy
REFERENCES Technologies (ICECET), Cape Town, South Africa, 2023, pp. 1-6, doi:
10.1109/ICECET58911.2023.10389513.
[1] W. F. Schneider and H. Guo, “Machine Learning,” The Journal of [16] “Network Slicing Recognition,” www.kaggle.com.
Physical Chemistry B, vol. 122, no. 4, pp. 1347–1347, Feb. 2018, doi: https://www.kaggle.com/datasets/gauravduttakiit/network-slicing-
10.1021/acs.jpcb.8b00035. recognition
[2] D. Bega, M. Gramaglia, A. Garcia-Saavedra, M. Fiore, A. Banchs, and X. [17] C. Angulo, X. Parra, and A. Català, “K-SVCR. A support vector machine
Costa-Perez, “Network Slicing Meets Artificial Intelligence: An AI- for multi-class classification,” Neurocomputing, vol. 55, no. 1–2, pp. 57–
Based Framework for Slice Management,” IEEE Communications 77, Sep. 2003, doi: 10.1016/s0925-2312(03)00435-1.
Magazine, vol. 58, no. 6, pp. 32–38, Jun. 2020, doi: [18] O.O. Shoewu, L.A. Akinyemi, L. A., and Edozie, “UAV Cellular
10.1109/mcom.001.1900653. Communication in 5G New Radio Wireless Standards,” In Unmanned
[3] A. Shalimov, V. Vishnevsky, and A. Gurtov, "Network slicing as a Aerial Vehicle Cellular Communications, pp. 25-45, Cham: Springer
service: Towards scalable network management in 5G," IEEE Network, International Publishing, 2022
vol. 34, no. 3, pp. 66-73, May/June 2020. [19] S.O. Oladejo, S.O. Ekwe, L.A Akinyemi, and S.A. Mirjalili, “The Deep
[4] X. Costa-Perez, M. Gramaglia, J. Alonso-Zarate, A. Banchs, X. Perez- Sleep Optimiser: A Human-Based Metaheuristic Approach, “IEEE
Costa, A. De Domenico, and A. Capone, "Network Slicing in 5G: A Access, 2023.
Survey on Principles, Enabling Technologies, and Solutions," in IEEE [20] A.T. Ajibare, S. O. Oladejo, S. O. Ekwe, L. A. Akinyemi and D.
Communications Surveys & Tutorials, vol. 21, no. 3, pp. 2369-2403, Ramotsoela, "Radiofrequency electromagnetic radiation exposure
Third Quarter 2019. DOI: 10.1109/COMST.2019.2910134. assessment, analysis, computation, and minimization technique in 5G
[5] I. Afolabi, T. Taleb, P. A. Frangoudis, M. Bagaa, and A. Ksentini, networks: A perspective on QoS trade-offs," in SAIEE Africa Research
‘‘Network slicing-based customization of 5G mobile services,’’ IEEE Journal, vol. 114, no. 4, pp. 114-127, Dec. 2023, doi:
Network., vol. 33, no. 5, pp.134–141, Sep. 2019. 10.23919/SAIEE.2023.10319380

Authorized licensed use limited to: University of Illinois. Downloaded on October 15,2024 at 14:04:20 UTC from IEEE Xplore. Restrictions apply.