A COMPARATIVE ANALYSIS OF MACHINE LEARNING ALGORITHMS ON NETWORK TRAFFIC FORECASTING

Abstract

Purpose- The purpose of this study is to evaluate and compare the performance of four different time series forecasting models applied to mobile network traffic data, a domain characterized by high variability and complex seasonal patterns. Accurate forecasting of mobile capacity needs in the telecommunications sector is of great importance for providing uninterrupted and high-quality service. Since each network is unique, it is necessary to build a model that best predicts the seasonal traffic changes of the network. Methodology- This research utilizes a comparative approach by implementing SARIMA, Prophet, LSTM, and a novel hybrid Prophet-LSTM model on monthly mobile traffic data from a telecommunications operator. The models were evaluated based on standard error metrics including MAE, MSE, RMSE and R2 score. Findings- The hybrid model leverages Prophet’s trend-seasonality decomposition with LSTM’s capability to learn nonlinear residual dynamics. The analysis reveals that the hybrid Prophet-LSTM model significantly outperforms the standalone SARIMA, LSTM, and Prophet models in terms of forecasting accuracy, flexibility, and adaptability. While SARIMA was limited in capturing complex, long-term trends and nonlinear fluctuations, LSTM required extensive hyperparameter tuning and was sensitive to data structure. Prophet proved to be effective in handling trend and seasonality with minimal parameter tuning, making it particularly suitable for cyclic patterns commonly observed in the telecommunications sector. However, the hybrid model’s ability to leverage Prophet’s decomposition strengths along with LSTM’s temporal learning capacity enabled it to deliver the most robust predictions with the lowest error rates. Conclusion- This model offers direct practical applications in network capacity planning, financial forecasting, and resource optimization processes. Moreover, it can be adapted for use in other sectors such as energy, transportation, and finance that rely heavily on time series data. Based on these findings, the hybrid Prophet-LSTM model is recommended for mobile traffic forecasting tasks involving both seasonal and nonlinear dynamics. Future studies may incorporate real-time streaming data and external factors to further improve predictive performance and real-world applicability.

Keywords

• Network traffic prediction
• SARIMA
• Prophet
• LSTM
• hybrid model

References

1. Aladag, C. H., Egrioglu, E., & Kadilar, C. (2009). Forecasting nonlinear time series with a hybrid methodology. Applied Mathematics Letters, 22(9), 1467–1470. https://doi.org/10.1016/j.aml.2009.02.006
2. Arslan, S. (2022). A hybrid forecasting model using LSTM and Prophet for energy consumption with decomposition of time series data. PeerJ Computer Science, 8, e1001. https://doi.org/10.7717/peerj-cs.1001
3. Bashir, S., Khan, A., Rehman, A., & Khan, M. (2022). Short-term electric load forecasting using hybrid Prophet–LSTM model. Energy Reports, 8, 345–355. https://doi.org/10.1016/j.egyr.2022.03.005
4. Cembaluk, P., Aniszewski, J., Knapińska, A., & Walkowiak, K. (2022). Forecasting the network traffic with PROPHET. Proceedings of the 3rd Polish Conference on Artificial Intelligence, 215–218.
5. Hyndman, R. J., & Athanasopoulos, G. (2018). Forecasting: Principles and practice (2nd ed.). OTexts. https://otexts.com/fpp2/
6. Karthika, S., Margaret, V., & Balaraman, K. (2017, August). Hybrid short term load forecasting using ARIMA-SVM. 2017 Innovations in Power and Advanced Computing Technologies (i-PACT) (pp. 1–6). IEEE. https://doi.org/10.1109/IPACT.2017.8245060
7. Katwal, A., Sharma, P., & Thakur, R. (2023). Comparative study of time series models for website traffic forecasting. Journal of Data Science Applications, 5(2), 44–53.
8. Khashei, M., & Bijari, M. (2011). A new class of hybrid models for time series forecasting. Expert Systems With Applications, 39(4), 4344–4357. https://doi.org/10.1016/j.eswa.2011.09.157

9. Kong, Y. H., Lim, K. Y., & Chin, W. Y. (2021). Forecasting Facebook user engagement using hybrid Prophet and long short-term memory model. International Conference on Digital Transformation and Applications (ICDXA), October 25-26, 2021, Kuala Lumpur, Malaysia. https://www.tarc.edu.my/files/icdxa/proceeding_new/920C21C7-4E28-44B4-9E3A-ECB64CBB13CD.pdf
10. Madan, R., & Mangipudi, P. S. (2018). Predicting computer network traffic: A time series forecasting approach using DWT, ARIMA and RNN. 2018 Eleventh International Conference on Contemporary Computing (IC3), 1–6. https://doi.org/10.1109/IC3.2018.8530608
11. Ning, Y., Kazemi, H., & Tahmasebi, P. (2022). A comparative machine learning study for time series oil production forecasting: ARIMA, LSTM, and Prophet. Computers & Geosciences, 164, 105126. https://doi.org/10.1016/j.cageo.2022.105126
12. Prajam, S., Wechtaisong, C., & Khan, A. A. (2022). Applying machine learning approaches for network traffic forecasting. Indian Journal of Computer Science and Engineering, 13(2), 324–330. https://doi.org/10.21817/indjcse/2022/v13i2/221302188
13. Shi, J., Leau, Y.-B., Li, K., & Obit, J. H. (2021). A comprehensive review on hybrid network traffic prediction model. International Journal of Electrical and Computer Engineering (IJECE), 11(2), 1450–1459. https://doi.org/10.11591/ijece.v11i2.pp1450-1459
14. Subashini, A., Sandhiya, K., Saranya, S., & Harsha, U. (2019). Forecasting Website Traffic Using Prophet Time Series Model. International Research Journal of Multidisciplinary Technovation, 1(1), 56–63. https://doi.org/10.34256/irjmt1917
15. Taylor, S. J., & Letham, B. (2017). Forecasting at scale. https://doi.org/10.7287/peerj.preprints.3190v2
16. Zaraket, K., Harb, H., Bennis, I., Jaber, A., & Abouaissa, A. (2024). Hyper-Flophet: A neural Prophet-based model for traffic flow forecasting in transportation systems. Simulation Modelling Practice and Theory, 134, 102954. https://doi.org/10.1016/j.simpat.2024.102954
17. Zhang, G. (2003). Time series forecasting using a hybrid ARIMA and neural network model. Neurocomputing, 50, 159–175. https://doi.org/10.1016/s0925-2312(01)00702-0