Load-Adjusted Transfer Learning for Limited Video-On-Demand Data

Video-On-Demand (VoD) systems face critical challenges in resource allocation when operating under data-constrained situations. Traditional Deep Learning (DL) methods for managing VoD networks depend on large datasets for training, which are frequently unattainable due to network disruptions, system failures, or inadequate data collection methods. To handle this, we propose a Transfer Learning Load Adjusted (TLLA) algorithm that enhances VoD systems' performance under limited training conditions. The TLLA transfers knowledge from pre-trained models by freezing sections of the neural network layers during retraining, reducing the need for large datasets. We freeze 50% (partial frozen) and 100% (fully frozen) of the pre-trained neural layers, and evaluate the model against fully trainable neural layers. Findings show that completely freezing neural layers achieves ≈40% of baseline performance, while partial neural layer freezing (50%) achieves ≈60% of baseline performance when measured using Root Mean Squared Error (RMSE) and R² metrics. These results demonstrate the success of transfer learning approaches in maintaining operability under severe freezing and limited VoD training data conditions. This study provides network managers and policy makers with actionable alternatives for monitoring Quality of Delivery (QoD) when input data is inadequate, enhancing robust resource allocation.