Multivariate Data Fusion-Based Learning of Video Content and Service Distribution for Cyber Physical Social Systems

TL;DR: This paper attempts to perform a multivariate analysis of video call record data collected from a wide area organizational network over a period of time and exhibits deviations from the conventional machine learning paradigms.

Abstract: Integration of physical processes with the computing world is driving newer challenges for networking frameworks. Cyber physical social systems (CPSSs) are another upcoming paradigm that encompasses the ever-growing interaction between the physical, social, and cyber worlds. As communication networks form the basis of these interactions, a cognitive evaluation of networks is called for. This CPSS driven network evolution was a direction motivating this paper. With the implementation of the next generation networks, traffic from real-time interactive services, such as video conferencing, is surpassing those of conventional transactional services. As such multimedia data transportation over IP networks has stringent quality constraints in terms of required bandwidth, latency, and jitter, legacy networks with no quality of service face challenges in terms of performance. We attempt to perform a multivariate analysis of video call record data collected from a wide area organizational network over a period of time. Learning-based prediction is attempted by training four classifiers: naive Bayes, $k$ -nearest neighbor, decision tree, and support vector machine. Two independent set of experiments were conducted with oversights of bandwidth and destination prediction. Both the discrete and continuous valued predictors were involved in the training. Performance evaluation of the generated hypothesis in both the cases was conducted using tenfold cross validation. Combined analysis using the assorted combinations of attributes was conducted, and thereafter, the effect of each feature was evaluated through singular attribute portioning. This paper presents observations, which exhibit deviations from the conventional machine learning paradigms. An attempt to increase the prediction accuracy of the classifiers was made through the boosting ensemble methodology. However, miniscule addition in performance was achieved. A maximum prediction accuracy of 81% for bandwidth and 60% for destination was obtained. Reasons of low accuracy of conventionally better performing algorithm were reasoned with a mathematical comprehension. Divergence of the obtained results from the accepted patterns poses an open research problem, particularly with respect to the nature and peculiarities of the data set. The proposed learning technique can have potential applications in social, tactical, and strategic spheres.