Multi-Task Learning at the Mobile Edge: an Effective Way to Combine Traffic Classification and Prediction

Mobile traffic classification and prediction are keytasks for network optimization. Most of the works in thisarea present two main drawbacks. First, they treat the twotasks separately, thus requiring high computational capabilities.Second, they perform data mining on the information collectedfrom the data plane, which is unsuitable for the mobile edge. Tobridge this gap, this paper properly tailors a Multi-Task Learningmodel running directly at the edge of the network to anticipateinformation on the type of traffic to be served and the resourceallocation pattern requested by each service during its execution.Our study exploits data mining from the control channel of anoperative mobile network to also reduce storage and monitoringprocessing. Different configurations of neural networks, whichadopt autoencoders (i.e. Undercomplete Autoencoder or Sequenceto Sequence Autoencoder) as key building blocks of the proposedMulti-Task Learning methodology for common feature repre-sentations, are investigated to evaluate the impact of the obser-vation window of traffic profiles on the classification accuracy,prediction loss, complexity, and convergence. The comparisonwith respect to conventional single-task learning approaches, thatdo not use autoencoders and tackle classification and predictiontasks separately, clearly demonstrates the effectiveness of theproposed Multi-Task Learning approach under different systemconfigurations