Feature vector generation for highly accurate traffic distribution prediction by supervised variational auto-encoder

Abstract

Traffic prediction is an important technique for network link capacity planning. Supervised Variational Auto-Encoder (SVAE), a deep learning technique, is a suitable approach for the network link capacity planning. The problem with the SVAE is that the mean absolute value error (MAPE) decreases when the correlation between the feature vector and the traffic probability distribution function (PDF) is low. In this paper, we propose an approach to increase the correlation by analytically obtaining values that correlate with the traffic PDF. Simulations are performed under the traffic conditions with the above problem to demonstrate the improvement in correlation and MAPE.