Scholarlink

<p>Cascade popularity prediction is a fundamental technique for information propagation analysis in many social applications such as Twitter, Facebook, Weibo, etc. Recent investigations show that the majority of information propagation (&gt; 61%) is contributed from numerous miniature cascades (cascade size &lt; 25). However, existing prediction methods are usually size-sensitive and have the basis for predicting large-size cascades. This causes seriously incorrect predictions for numerous miniature cascade predictions. This is because miniature cascade suffers a <i>feature starvation</i> problem while large-size cascade has much more passing features. In this paper, we propose a novel cascade feature augmentation method, named CasMV, which introduces additional mutual information from a multi-view of the cascade. In detail, the first view is the cascade graph at the observing time. The second view is the cascade graph during the passing process. The last view is the user graph across the different cascades. After, a temporal fusion network is proposed to fuse the cascade features from the above views. At last, the fused cascade features are used to predict the future propagation size of each cascade. Experiments on real-world social graphs show that the proposed method achieves new state-of-the-art records under the scenario where miniature cascades dominate the data, obtaining approximately 5% improvements compared to peer methods. Moreover, the proposed method achieves competitive results in the general data environment compared to the SOTA methods. The source code of the work is available at <a href="https://github.com/subetter/CasMV.git">https://github.com/subetter/CasMV.git</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

Multi-view temporal graph neural network for numerous miniature cascade popularity prediction

  • 期刊论文

  • 发表日期：

    2025年6月9日

查看全文
  • Yasu Wu,
  • Changlong Fu,
  • Zhenli He,
  • Wei Huang,
  • Cheng Xie

摘要

Cascade popularity prediction is a fundamental technique for information propagation analysis in many social applications such as Twitter, Facebook, Weibo, etc. Recent investigations show that the majority of information propagation (> 61%) is contributed from numerous miniature cascades (cascade size < 25). However, existing prediction methods are usually size-sensitive and have the basis for predicting large-size cascades. This causes seriously incorrect predictions for numerous miniature cascade predictions. This is because miniature cascade suffers a feature starvation problem while large-size cascade has much more passing features. In this paper, we propose a novel cascade feature augmentation method, named CasMV, which introduces additional mutual information from a multi-view of the cascade. In detail, the first view is the cascade graph at the observing time. The second view is the cascade graph during the passing process. The last view is the user graph across the different cascades. After, a temporal fusion network is proposed to fuse the cascade features from the above views. At last, the fused cascade features are used to predict the future propagation size of each cascade. Experiments on real-world social graphs show that the proposed method achieves new state-of-the-art records under the scenario where miniature cascades dominate the data, obtaining approximately 5% improvements compared to peer methods. Moreover, the proposed method achieves competitive results in the general data environment compared to the SOTA methods. The source code of the work is available at https://github.com/subetter/CasMV.git.