Abstract:
Today’s Internet is a combination of networks of different kinds and sizes, supporting a wide
range of applications that may vary in quality of service (QoS) requirements. Network
congestion problem has a significant influence on those applications QoS requirements. So,
learning-based traffic flow identification and congestion problem management are becoming key
components in computer networks since they are critical issues. For decades, rule-based
congestion control algorithms have been created and applied to control those concerns using
established rules. In rule-based congestion management, the network response or feedback
information is mainly used as an input to handle those issues. But it might lead to a
misunderstanding of the congestion signal, and it has an impact on the QoS requirements of the
end-user’s application since it ignores their resource requirements. The application's QoS
requirements, mainly bandwidth and latency requirements, provide some extra guidance to the
network controller in selecting an efficient controlling algorithm to preserve the network from
congestion issues.
This paper focuses on developing a learning-based technique that regulates the rate at which
hosts send packets into the network without affecting their application's QoS requirements with
the help of an ML-powered prediction system. We employed an ensemble of five conventional
machine learning (ML) methods to address the problem and boost prediction accuracy. After
basic data mining tasks, the base-level model and metamodel were trained and validated using
ISCXTor 2016, a prominent and current benchmark dataset. In this work, a 70%–30% split
method was used. In this technique, 70% of the whole dataset is used to train the estimator, and
the remaining 30% is used to test the model. Parameter tuning, feature selection, and cross validation (CV) principles were used, and a high accuracy result of 88.33% was obtained. A
maximum cross-validated accuracy of 88.89% was obtained.
