Enhancing QoE Prediction for Video Streaming Using a Hybrid CNN-LSTM Model

Quality of Experience (QoE) prediction in video streaming plays an important role in improving user satisfaction and optimizing adaptive strategies. The current nonhybrid QoE prediction models tend to suffer from poor accuracy due to the inability to model complex and varied temporal/spatial dependencies. We suggest using a hybrid model that combines convolutional neural networks (CNNs) to pick up on visual details and long short-term memory (LSTM) networks to understand time-based patterns. This combined method works well for estimating QoE by considering both the visual details and the timing of video data, even when network conditions change. To do this research, we applied this set of metrics to the Waterloo SQoE-IV dataset, a robust benchmark for QoE prediction. We train and evaluate the proposed model under various streaming conditions, taking into account network fluctuations, adaptive bitrate streaming strategy, and user behavior. To improve the predictive performance of our framework, feature selection methods are incorporated to remove redundancy, reduce computational complexity, and retain high accuracy. The experimental results showed that the CNN-LSTM model is much better than the current QoE prediction methods, achieving a linear correlation coefficient (LCC) of 0.8929 and a Spearman rank-order correlation coefficient (SROCC) of 0.8846. We also dramatically reduced prediction errors (RMSE) to 7.7323 subjective QoE scores. This evidence suggests the efficacy of our approach to grab both the spatial and the temporal dependencies, which makes the QoE prediction more robust. Therefore, this work helps develop intelligent streaming adaptation mechanisms to enhance and smooth video streaming services for end users.