LA-CNN: Load-Adjusted Video-on-Demand Prediction using CNNs

The ability to predict RTP packet counts accurately is needed if network managers are to be able to manage Video-on-Demand (VoD) sessions. We contribute a new algorithm called Load-Adjusted Convolutional Neural Networks (LA-CNNs) which addresses the task of accurately predicting the number of RTP packets received by a VoD client. The objective of this paper is to evaluate the performance of LA-CNN and to compare it with an Un-Adjusted CNN (UA-CNN) algorithm and a set of other classical benchmark algorithms, which include Elastic Net (EN), Ridge Regression (RR) and the Least Absolute Shrinkage and Selection Operator (LASSO) in both UA and LA forms. We find that LA-CNN and UA-CNN give 20% better performance than LA and UA (EN, RR, LASSO) when the Root Mean Squared Error (RMSE) and (R2) are measured. Moreover, the LA-CNN gives 35% performance gain over UA-CNN when RTP packet count predictions are compared. These findings are important in the context of network administration and management as they provide evidence that Load-Adjusted learning provides consistent performance gains when a CNN is used as the learning algorithm.