Understanding the Effect of Opinion Polarization in Short Video Browsing

TL;DR

This work tackles Opinion Polarization (OP) in short video browsing by leveraging EEG-based neural signals to quantify the neural processing of polarized content and to predict exposure to polarized videos. Using a three-stage, multimodal user study with 24 participants and 230 user–personage pairs, the authors demonstrate that OP significantly shifts explicit sentiment judgments and modulates EEG patterns, particularly in the $\delta$ and $\gamma$ bands, with engagement factors (likes, viewing time) shaping the strength of neural correlates. Explicit sentiment signals outperform traditional implicit behavior signals (likes, dwell time) for capturing OP effects, while EEG contributes strong predictive power, especially when fused with explicit sentiments and behavior. The study introduces a binary OP-detection framework based on $0/1$ labels and demonstrates that combining EEG with sentiment and behavior features yields the best $AUC$, $ACC$, and $F1$ performance, suggesting practical potential for real-time OP monitoring and content recommendation adjustments. Overall, the work provides a novel, EEG-informed perspective on OP in AI-driven information channels, with implications for designing healthier short-video ecosystems and informing future research on neural signals in information retrieval.

Abstract

This paper explores the impact of Opinion Polarization (OP) in the increasingly prevalent context of short video browsing, a dominant medium in the contemporary digital landscape with significant influence on public opinion and social dynamics. We investigate the effects of OP on user perceptions and behaviors in short video consumption, and find that traditional user feedback signals, such as like and browsing duration, are not suitable for detecting and measuring OP. Recognizing this problem, our study employs Electroencephalogram (EEG) signals as a novel, noninvasive approach to assess the neural processing of perception and cognition related to OP. Our user study reveals that OP notably affects users' sentiments, resulting in measurable changes in brain signals. Furthermore, we demonstrate the potential of using EEG signals to predict users' exposure to polarized short video content. By exploring the relationships between OP, brain signals, and user behavior, our research offers a novel perspective in understanding the dynamics of short video browsing and proposes an innovative method for quantifying the impact of OP in this context.

Figures (8)

• Figure 1: The overall procedure of the three-stage user studies. $S_1$: Participants browsed videos featuring positive or negative polarity on these 10 personages. Their behavior signals, brain signals, and explicit responses are collected during the video browsing process or the post-experiment questionnaire. $S_2$: Participants browsed short videos about these personages posted by the Platform in a field study lasting 6 days. For each personage, the Platform only recommended videos with either positive or negative polarity, which acts as a manifestation of OP. $S_3$: The last stage resembles the procedures in stage $S_1$. However, it diverges in the selected videos and the timing of the questionnaire.
• Figure 2: Webpage platform playback page.
• Figure 3: Sentiment scores averaged across participants and personages in pre- and post-OP environments.
• Figure 4: Difference of sentiment score collected pre- and post-environment for different personages and distractors, averaged across participants. CHC, TE, ..., and MC are the abbreviations of the ten selected personages. Distractor indicates the average score across all distractors for the experimental control.
• Figure 5: Averaged like rate for videos with positive (Pos-videos) or negative opinions (Neg-videos).