嘉    宾：肖帅勇 副教授 同济大学
主持人：韩雪雯 助理教授 上海交通大学安泰经济与管理学院
时    间：2026年4月23日（周四）13:30-15:00
地    点：上海交通大学徐汇校区安泰楼A507

内容简介:
Predicting the top-viewed clips within a video offers a promising solution for video creators and platforms to pop their videos, thereby driving more effective traffic monetization and generating higher commercial profits. Top-viewed video clip prediction involves fine-grained viewership prediction (FVP), i.e., predicting the viewership for each clip and then selecting the top-viewed ones. Guided by the signal detection theory, we design a novel deep learning method named highlight detection-enhanced FVP (HD-FVP), synergizing three innovative components: (1) clip-wise multimodal discrepancy learning, which detects highlight clips through differentiating the multimodal representations of informative clips from those of regular ones, (2) adaptive in-context learning, which incorporates the contextual affinity among adjacent clips in predicting clip-wise viewership, and (3) highlight-enhanced clip-wise viewership prediction, which coordinates the refined multimodal representations of clips with the estimated highlight scores of the clips to further enhance the clip-wise viewership prediction. Empirical evaluation based on data from two prominent online video platforms demonstrates the superiority of HD-FVP over state-of-the-art benchmarks. Exploratory analyses and economic value analyses render insights into how highlight clips function differently from regular ones in FVP and the commercial potential of HD-FVP.

演讲人简介：
肖帅勇，同济大学经济与管理学院副教授。研究方向聚焦于短视频营销等商业应用场景，以管理学/统计学理论为驱动，致力于通过深度学习技术创新，提升人工智能在多模态用户行为感知、认知与推理方面的能力和应用。研究成果发表于MIS Quarterly 、Information Systems Research、 INFORMS Journal on Computing等期刊。主持国家自然科学基金青年项目、科技部重点研发计划、博士后科学基金面上资助项目等多项研究课题。

欢迎广大师生参加！