Scholarlink

Effective popularity forecasting in time series multimedia data is necessary for the safety of social network and cyberspace. Nevertheless, establishing an accurate time series forecasting system as well as analyzing user distribution and user interactions is a gap that needs to be filled. This study proposes KanPaTST, an innovative patch time series forecasting method which incorporates multiple levels of correlative data and users’ Myers-Briggs Type Indicator (MBTI) classifications based on contextual token processing. The method employs transformer on representation pre-training and Kan network on multi-modal informed data and trend feature fine-tuning to reach an accurate public opinion popularity forecasting. By leveraging the capabilities of transformer networks to infer the broader temporal trends and MBTI distribution classifier to capture users’ MBTI map spatial features, our approach effectively forecast popularity within both trend threshold and forecasting threshold. Extensive experiments on three real-world public datasets demonstrate that KanPaTST can significantly outperform most baseline methods in both long-term and short-term forecasting. Specifically, KanPaTST enhances MSE and MAE scores by a average of 39.3% and 16.1% while maintaining stability under shifted data joint training. Furthermore, the ablation study furnishes empirical evidence supporting the effectiveness of each component with KanPaTST.

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

Trend Prediction First, Personality Refinement After. KanPaTST: A KAN Fine-Tuned Patch Time Series Transformer for Public Opinion Popularity Forecasting

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Shaoguo Cui,
  • Sifan Zhao,
  • Linfeng Gong,
  • Wei Xia,
  • Yulong Yang,
  • Binbin Sang,
  • Tiansong Li,
  • Xiaoxin Chen

摘要

Effective popularity forecasting in time series multimedia data is necessary for the safety of social network and cyberspace. Nevertheless, establishing an accurate time series forecasting system as well as analyzing user distribution and user interactions is a gap that needs to be filled. This study proposes KanPaTST, an innovative patch time series forecasting method which incorporates multiple levels of correlative data and users’ Myers-Briggs Type Indicator (MBTI) classifications based on contextual token processing. The method employs transformer on representation pre-training and Kan network on multi-modal informed data and trend feature fine-tuning to reach an accurate public opinion popularity forecasting. By leveraging the capabilities of transformer networks to infer the broader temporal trends and MBTI distribution classifier to capture users’ MBTI map spatial features, our approach effectively forecast popularity within both trend threshold and forecasting threshold. Extensive experiments on three real-world public datasets demonstrate that KanPaTST can significantly outperform most baseline methods in both long-term and short-term forecasting. Specifically, KanPaTST enhances MSE and MAE scores by a average of 39.3% and 16.1% while maintaining stability under shifted data joint training. Furthermore, the ablation study furnishes empirical evidence supporting the effectiveness of each component with KanPaTST.