Revisiting the Relation between QoS and QoE to improve predictive Models in Adaptive Streaming

The rapid adoption of video streaming services has significantly transformed media consumption behavior. Increasingly, users are choosing these platforms over traditional TV broadcasts. However, this surge in demand poses challenges for content providers in maintaining high-quality service due to limited transmission resources and conflicting interests between users and providers. Users strive for a high Quality of Experience (QoE), often associated with a high Quality of Service (QoS), particularly a high bitrate. However, content providers face the challenge of accommodating an increasing number of concurrent users. To achieve this, content providers may need to limit the streaming bitrate. Various methods, including machine learning models, have been developed to infer the best trade-off between user QoE needs and provider scalability requirements. These approaches require the building of accurate prediction/inference QoE models, which is still an open challenge. This is especially true for subjective QoE metrics such as user engagement. Despite this, engagement modeling is gaining attention due to its explainability. To improve engagement prediction accuracy, this paper proposes a model that explores the correlation between QoE and QoS metrics, and client bitrate switching decisions. Combining these metrics increased the accuracy of engagement prediction by 10% compared to models using traditional QoS metrics of the application. Using this prediction model, we observed a significant increase in user potential engagement, with an average gain of 100%. Furthermore, average bandwidth savings reached several gigabytes, but the impact on QoE for users due to this reduction was minimal, less than 0. 4% compared to the original QoE.