Video-on-Demand Prediction via Ensemble Load-Adjusted CNN-Random Forests

Using accurate RTP packet count predictors can increase the resilience of network resources in Real-Time Trans-port Protocol (RTP) Video-on-Demand (VoD) systems. We in-vestigate RTP packet count prediction in VoD environments by introducing a stacked ensemble Load Adjusted Convolutional Neural Network and Random Forest (LA-CNN-RF) predictor. We explore the performance of the LA-CNN-RF predictor and compare it with an Un-Adjusted predictor called UA-CNN-RF and the state-of-the-art LA-CNN and UA-CNN models. We compute the Root Mean Squared Error (RMSE) and R2 score of the resulting predictions from LA-CNN-RF and introduce a threshold to assess the RTP prediction improvement rate of CNN-RF when we consider both the Load Adjusted (LA) and Un-Adjusted (UA) CNN variants. Findings show that LA-CNN-RF improves the RTP packet count predictions by ≈35 % while LA-CNN-RF outperforms UA-CNN-RF by ≈15%. These findings are important for network administrators looking for strategies to achieve a consistent network state by improving network traffic management systems and enhancing predictor capabilities.