Prediphant: Short Term Heavy User Prediction

Abstract

Traffic prediction is of paramount importance for the correct management of network infrastructures. Most research efforts try to forecast the aggregated traffic over the network and over large time windows. In this work, we tackle the problem the other way around. That is, we predict the behaviour of individual users over short time windows. First, we investigate the contribution of the most data eager users to the global network traffic. We do it by analyzing network traces coming from several thousand real users. Then, we design a technique, based on a combination of Natural Language Processing (NLP) and Long Short-Term Memory (LSTM) machine learning models, that leverages past navigation patterns to predict sudden changes in the amount of resources consumed by each user. Finally, we evaluate our method using real data finding it is able to predict about 80% of the users that will rump up their network needs in most realistic scenarios.