Cognitive prediction of end-to-end bandwidth utilisation in a non-QoS video conference

The most sought after bandwidth killer application on networks has been video conference. For a complex network, specifically based within an organization, implementing quality of service (QoS) is administratively not always feasible as the priorities are regulated. Predicting future network traffic in a non-QoS implemented network by using information about the source, destination and application can give preparatory time to make the network ready for unstable and random demands. The paper uses machine learning techniques to predict the bandwidth utilization of an end-to-end video conference session. Experimental results in this paper show that these features work well in detecting the bandwidth utilization. These experiments were done on a corpus of 24,000 video conference connections. The cognition is based on experimenting on features such as time of call, source, destination, call type, expected duration and cause codes. The result is based on combination of all these features which gave an accuracy of more than 78% on real traffic using two of the common classifiers - k-nearest neighbors and tree based classifier. Support vector machine (SVM) and Naive Bayes gave lower learning accuracy. Prediction results were also obtained by varying the combination of features to detect the predominating features in the cognition. It has been established that the bandwidth at which the connection is established is not entirely dependent on the source and destination but the other features also play a role in deciding the bandwidth of the connection. The prediction accuracy further increases if video calls are allowed only at discrete pre-designated bandwidth levels.