Network Traffic Prediction using Quantile Regression with linear, Tree, and Deep Learning Models

Abstract

Machine Learning research has progressed tremendously in recent years. Major fields that machine learning pushed its frontier were prediction and data modeling. In this work we evaluate the applicability of a handpicked prediction models on predicting inter-day aggregate network traffic. We chose models that work best with multi-variate feature space. They represent linear, decision trees, and neural network models. Over the years, predicting network traffic has resorted to predicting point values. This approach is not descriptive enough and naively gives a shallow conclusion about the data. We propose using a quantile loss function that predicts boundaries or prediction intervals. Our results show that linear models fared well compared to their simplicity while Long Short-Term Memory Neural Networks gave best results across all experiments.