Scientific Publications Directory

Full Author List

Purnawansyah, Department of Information Systems, Universitas Muslim Indonesia, Indonesia; Wibawa, Aji Prasetya; Widiyaningtyas, Triyanna Department of Electrical Engineering and Informatics, Universitas Negeri Malang, Indonesia; Haviluddin, Department of Informatics, Universitas Mulawarman, Indonesia; Raja, Roesman Ridwan; Darwis, Herdianti, Department of Informatics, Universitas Muslim Indonesia, Indonesia; Nafalski, Andrew, UniSA Education Futures, School of Engineering, University of South Australia, Australia

Abstract

Network congestion arises from factors like bandwidth misallocation and increased node density leading to issues such as reduced packet delivery ratios and energy efficiency, increased packet loss and delay, and diminished Quality of Service and Quality of Experience. This study highlights the potential of deep learning and ensemble learning for network congestion analysis, which has been less explored compared to packet-loss based, delay-based, hybrid-based, and machine learning approaches, offering opportunities for advancement through parameter tuning, data labeling, architecture simulation, and activation function experiments, despite challenges posed by the scarcity of labeled data due to the high costs, time, computational resources, and human effort required for labeling. In this paper, we investigate network congestion prediction using deep learning and observe the results individually, as well as analyze ensemble learning outcomes using majority voting, from data that we recorded and clustered using K-Means. We leverage deep learning models including BPNN, CNN, LSTM, and hybrid LSTM-CNN architectures on 12 scenarios formed out of the combination of level datasets, normalization techniques, and number of recommended clusters and the results reveal that ensemble methods, particularly those integrating LSTM and CNN models (LSTM-CNN), consistently outperform individual deep learning models, demonstrating higher accuracy and stability across diverse datasets. Besides that, it is preferably recommended to use the QoS level dataset and the combinations of 3 clusters due to the most consistent evaluation results across different configurations and normalization strategies. The ensemble learning evaluation results show consistently high performance across various metrics, with accuracy, Matthews Correlation Coefficient, and Cohen's Kappa values nearing 100%, indicates excellent predictive capability and agreement. Hamming Loss remains minimal highlighting the low misclassification rates. Notably, this study advances predictive modeling in network management, offering strategies to enhance network efficiency and reliability amidst escalating traffic demands for more sustainable network operations.

Related SDGs Targets