Scholarlink

<p>User-generated videos (UGVs) have become a dominant form of online content on social media. Predicting popular UGVs holds substantial commercial value in fields such as advertising, edge computing, and recommender systems. However, this research faces the following challenges: (1) UGVs can be described by a variety of heterogeneous content features; (2) capturing popularity trends requires consideration of their mixed patterns; (3) the influence of content and popularity trend features on popularity evolution varies across different time periods. In this paper, we propose a Temporal Information-aware Multimodal Learning network (TIML) to address these challenges, based on data collected from the leading UGV platform Bilibili. First, by analyzing the evolution patterns of UGV popularity, we extract four types of temporal context features: non-workday eve, non-workday, published day count, and time period. Second, we extract a set of multimodal content features–including visual, acoustic, textual, social, and metadata—to represent UGV static attributes. Third, we propose a Hierarchical Popularity Trend Feature Extraction network (HPTFE) as the temporal modeling module to effectively capture popularity trends. Finally, we design a Time-aware Attention Feature Fusion (TAFF) method to adaptively fuse content and popularity trend features by modulating the attention mechanism with temporal context features. Our proposed TIML significantly outperforms baseline methods on two real-world datasets, with improvements of 17.87% and 13.23% in MSE. Code and dataset are available at <a href="https://github.com/RolyPo1y/TIML">https://github.com/RolyPo1y/TIML</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Temporal information-aware multimodal learning network for user-generated video popularity prediction

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Mingjun Xi,
  • Pingshan Liu,
  • Seshu Yu

摘要

User-generated videos (UGVs) have become a dominant form of online content on social media. Predicting popular UGVs holds substantial commercial value in fields such as advertising, edge computing, and recommender systems. However, this research faces the following challenges: (1) UGVs can be described by a variety of heterogeneous content features; (2) capturing popularity trends requires consideration of their mixed patterns; (3) the influence of content and popularity trend features on popularity evolution varies across different time periods. In this paper, we propose a Temporal Information-aware Multimodal Learning network (TIML) to address these challenges, based on data collected from the leading UGV platform Bilibili. First, by analyzing the evolution patterns of UGV popularity, we extract four types of temporal context features: non-workday eve, non-workday, published day count, and time period. Second, we extract a set of multimodal content features–including visual, acoustic, textual, social, and metadata—to represent UGV static attributes. Third, we propose a Hierarchical Popularity Trend Feature Extraction network (HPTFE) as the temporal modeling module to effectively capture popularity trends. Finally, we design a Time-aware Attention Feature Fusion (TAFF) method to adaptively fuse content and popularity trend features by modulating the attention mechanism with temporal context features. Our proposed TIML significantly outperforms baseline methods on two real-world datasets, with improvements of 17.87% and 13.23% in MSE. Code and dataset are available at https://github.com/RolyPo1y/TIML.