Abstract
Emerging video services introduce functional parameters that directly influence user cognitive states and experience quality, termed Functional Quality of Experience (fQoE). Electroencephalogram (EEG) signals provide rich information about brain activity, revealing patterns in brain networks that can serve as a reliable basis for fQoE assessment. This paper proposes a novel EEG-based brain network construction method to evaluate fQoE and investigate its underlying neural mechanisms. By manipulating functional parameters to induce varying fQoE levels, we collected EEG data and extracted multi-electrode features fused into graph structures. A self-attention graph pooling model was then used to identify fQoE-sensitive brain networks. Using bullet comment (danmaku) coverage as a case study, our method achieved classification accuracies between 80% and 86% across various video types. Results indicate that higher fQoE is associated with reduced brain connectivity in specific regions, providing a quantitative tool for optimizing video service parameters.
Introduction
Traditional Quality of Experience (QoE) research has primarily focused on technical transmission factors such as network latency, jitter, and video bitrate, often referred to as technical QoE (tQoE). However, modern video platforms offer a variety of functional features—such as interactive "bullet comments" (danmaku), virtual avatars, and personalized interfaces—that significantly impact user satisfaction. This dimension of user experience is defined as Functional Quality of Experience (fQoE).
While fQoE is critical for service optimization, it is traditionally measured through subjective surveys, which are prone to bias and lack physiological depth. This research addresses the need for an objective, neurophysiologically grounded assessment method by leveraging EEG signals to model the brain's response to functional video parameters.
Methodology
The proposed framework transforms raw neural activity into an interpretable assessment of user experience through a multi-stage pipeline.
EEG Data Preprocessing
To ensure high-quality signal analysis, raw EEG data undergoes rigorous preprocessing. This includes removing unusable electrodes, filtering power line interference, and applying Independent Component Analysis (ICA) to eliminate artifacts such as eye blinks. The continuous signal is then segmented using a sliding window approach to capture temporal dynamics.
Brain Network Construction
The core of the method lies in modeling the brain as a graph $G = (V, E)$.
- Nodes ($V$): Represent individual EEG electrodes. Node features are derived from the Power Spectral Density (PSD) across five frequency bands ($\delta, \theta, \alpha, \beta, \gamma$).
- Edges ($E$): Represent functional connectivity between brain regions, calculated using Pearson correlation coefficients.
A Graph Convolutional Network (GCN) is employed to capture the spatial topology of these networks. A key innovation is the Self-Attention Graph Pooling (SAGPool) mechanism, which automatically identifies the most relevant brain regions for fQoE classification, enhancing both performance and interpretability.
Experimental Design and Dataset
The study utilized "Bullet Chatting Videos" (BCV) from Bilibili to induce different fQoE levels. Fifty healthy participants watched videos across five categories (Auto-tune, Technology, Dance, Film, and Music) while their EEG signals were recorded using a 128-lead system.
The primary functional parameter manipulated was the danmaku coverage rate, set at five levels: 0%, 25%, 50%, 75%, and 100%. This setup allowed researchers to observe how varying levels of visual clutter and information density directly affect neural patterns.
Results and Discussion
The proposed model demonstrated superior performance in classifying fQoE levels compared to standard benchmarks like EEGNet.
Classification Performance
The model achieved high accuracy across all video types:
- Auto-tune (Guichu): 86%
- Music: 84%
- Film: 82%
- Technology: 81%
- Dance: 80%
Neurophysiological Insights
The analysis revealed that higher fQoE (better experience) is correlated with a decrease in functional connectivity in the frontal pole, middle frontal gyrus, parietal, and temporal lobes. This suggests that excessive functional elements, like dense bullet comments, increase cognitive load, forcing the brain to increase its connectivity to process competing visual and linguistic information. Conversely, a "cleaner" and more satisfying experience allows for more efficient neural processing with fewer active connections.
Furthermore, features from the $\beta$ and $ \gamma$ bands were found to be the most discriminative, confirming that fQoE involves high-level cognitive processes such as active concentration and emotional integration.
Conclusion
This research provides the first EEG-based brain network construction method specifically designed for fQoE assessment. By identifying the neural markers of functional satisfaction, the method offers a quantitative, objective tool for video service providers to optimize their functional parameters. Future work will focus on real-time assessment and improving model robustness across diverse user populations.
Summary updated