community resonance prediction

This work addresses the limitations of traditional video quality assessment, which emphasizes visual fidelity yet fails to capture the community engagement and resonance elicited by user-generated content (UGC). To bridge this gap, the paper introduces a novel task, CASTER, which evaluates whether UGC elicits positive community feedback through multimodal attribute analysis, and proposes the MEDEA architecture to enable human-centered quality judgment. The core innovations include the first-of-its-kind Social Chain-of-Thought mechanism that simulates diverse viewer perspectives to model “community mind,” and a hybrid training strategy combining supervised fine-tuning with process-based reinforcement learning guided by a Social Alignment Reward to align reasoning pathways with human social cognition. Evaluated on the newly curated human-annotated benchmark CASTER-Bench, the proposed approach substantially outperforms existing models, yielding interpretable reasoning that closely aligns with real-world community responses.

Existing studies predominantly extract features directly from narrative content to predict user engagement, neglecting the modeling of audience forward-looking expectations—i.e., readers’ beliefs about future story developments. This work introduces the first generative framework grounded in large language models (LLMs), which explicitly characterizes prospective psychological variables—including reader expectations, uncertainty, and surprise—via multi-path narrative continuation. These variables are systematically integrated into engagement prediction. To our knowledge, this is the first approach enabling computationally tractable modeling of forward-looking beliefs over unstructured narrative data. Evaluated on over 30,000 novel chapters, the method achieves an average 31% gain in marginal explanatory power over conventional feature engineering, significantly improving predictive performance across diverse engagement behaviors—including reading duration, commenting, and voting.

Understanding what content is valued across different online communities is crucial for social computing tasks such as recommendation, moderation, and ranking. This work proposes VASTU, the first standardized, cross-community benchmark enabling methodologically comparable evaluation, comprising 75,000 comments from 15 Reddit communities along with community endorsement labels and linguistic features. Through systematic evaluation of feature-engineered models, Transformers, and large language models (LLMs)—including both prompting and fine-tuning approaches—under both global and community-specific paradigms, the study finds that community-specific models substantially outperform generic ones. Fine-tuned Transformers achieve the best performance (AUROC = 0.72), with smaller fine-tuned models (0.65) surpassing prompted LLMs (0.60), while reasoning-based models perform worst (0.53), challenging assumptions about the efficacy of LLM reasoning for this task.

This study investigates the disruptive impact of generative AI (AIGC) on self-governance in online communities, focusing on three key risks faced by Reddit moderators: declining content quality, erosion of community trust, and imbalanced rule enforcement. Through 15 semi-structured in-depth interviews with active moderators, analyzed via thematic coding and grounded theory, the research systematically uncovers AIGC’s profound effects on communal social value and collaborative normative structures—the first such empirical investigation. Findings reveal that current technical detection methods are inefficient and time-intensive, whereas community-driven norm co-creation practices significantly mitigate AIGC-related disruptions. Accordingly, the study proposes a novel governance paradigm—“human-centered co-governance over technological dependency”—which reorients moderation toward participatory, values-aligned collective action. This framework offers a theoretically grounded and empirically informed, locally contextualized pathway for platform governance and AI ethics, bridging critical gaps between algorithmic affordances and sociotechnical sustainability.

This study addresses the problem of predicting user engagement (e.g., like ratio, comment sentiment) for long-form YouTube videos and quantifying the relative contributions of textual, audio, and visual modalities. To this end, we propose an interpretable multimodal attention-based deep learning framework that integrates self-attention with cross-modal feature interaction, and introduce a novel attention-guided posterior pruning technique to eliminate spurious statistical associations. Empirical results across multiple real-world datasets demonstrate that the textual modality contributes most substantially to engagement prediction; within the first 30 seconds of videos, auditory cues predominantly drive linguistic interaction sentiment, whereas visual cues govern non-linguistic interaction tendencies. The model exhibits strong generalization performance and yields theoretically grounded, actionable marketing insights—such as modality-specific timing effects on viewer behavior—thereby advancing both methodological rigor and practical applicability in multimodal engagement modeling.

Short-form videos pose significant challenges for standardized modeling of user engagement due to their multimodal content and platform-specific algorithms. This study addresses this gap by computationally operationalizing classical interpretive theories from narratology, rhetoric, communication studies, and semiotics at scale. Leveraging a multimodal large language model, we automatically annotated 77 theory-driven structural variables across approximately 10,000 TikTok videos from Estonian brands and institutions, supplemented by human validation to assess reliability. Controlling for account size and video age, our model yielded a stable, albeit modest, improvement in predicting user engagement. Results indicate that variables related to perception and communication were reliably annotated, whereas deeper semiotic and archetypal structures proved more challenging to capture. This work establishes a systematic computational framework for analyzing the cultural structures embedded in short-form video content.

Existing research often examines textual, visual, or audio modalities of short videos in isolation, failing to uncover how their interplay influences user engagement. This work proposes the first reproducible and interpretable multimodal analysis framework that integrates automated feature extraction with Shapley-value-based attribution to systematically investigate how multimodal interactions affect view counts in TikTok content related to social anxiety disorder. The study reveals that facial expressions are more predictive than textual sentiment, that informational content garners greater attention than emotional support, and that multimodal synergies exhibit strong threshold-dependent effects—thereby transcending the limitations of conventional unimodal analyses.