Estimating Video Quality Using Coarse-Grained Features: Insights and Limitations from Gaussian Mixture Models

The significant demand for high-quality video streaming, and the wide proliferation of this application in entertainment, education and communication, has created an urgent need for methods that ensure user Quality of Experience (QoE). Traditional approaches for predicting video playback quality have focused on fine-grained network layer features, which can be computationally intensive and require extensive data collection. To reduce the high computation demand and data collection requirements, we propose to use machine learning methods to predict mean video playback bitrate using coarse-grained features derived from real-time packet scheduling priority weights. Although coarse-grained features alone may not always perform optimally due to data ambiguity, they can still enable meaningful predictions in specific contexts. In this work, we investigate the use of coarse-grained features and their ability to predict video streaming QoE. To do so, we introduce a Gaussian Mixture Model (GMM)-based filtering technique to identify regions where the model performs well. Our evaluation shows that while our model initially achieves a 56% macro-average F1 score for the entire dataset, it reaches an 81% macro-average F1 score with the filtered subset of data. This approach highlights the potential of coarse-grained features for accurate QoE estimation by identifying the regions of the dataset where these features possess sufficient discriminative power, thereby enhancing domain knowledge and the trustworthiness of model outputs.