| Video-on-Demand (VoD) networks typically perform better when appropriate Machine Learning (ML) models are used in data constrained scenarios. We present a Multi-Layered Transfer Learning (MLTL) algorithm for addressing missing data in VoD experiments. MLTL utilizes self-transfer learning by adding an extra layer on top of a trained neural layer prior to final training. MLTL is compared with the Uni-Layered Transfer Learning (ULTL) approach, in which the TL model’s neural layers are not stacked prior to training. We conduct a performance analysis of both MLTL and ULTL on a time-varying load dataset used in the State-of-the-art Transfer Learning Load Adjusted (TLLA) experiments. We apply MLTL and ULTL to 1) train the ML model from scratch; the initial dataset before a pre-trained model is designed, and 2) with 50% freezing of the neural layers. We assess the performance of both algorithms using Root Mean Squared Error (RMSE) when the number of concurrent users of the system is in the range 20 ≤ k ≤ 80. The results reveal cases where the training-from-scratch models outperform the ULTL models, whereas MLTL outperforms all training-from-scratch instances. The findings motivate the application of MLTL in VoD systems when training data is limited due to factors such as dynamically changing network conditions |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.