Scholarlink

<p>Bayesian personalized ranking (BPR) has gained prominence as an effective method for pairwise learning, particularly in personalized tasks such as recommendation systems. However, recent developments in adversarial machine learning (AML) have raised concerns about the vulnerability of advanced BPR techniques. While adversarial training has proven effective in enhancing resilience and performance in the recommendation system (RS) domain, its application in network embedding remains underexplored. This study addresses the gap by introducing the adversarial neural Bayesian ranking-based academic network embedding (Adversarial Neural-Brane). The proposed approach integrates a neural network model with BPR to generate latent representations of vertices, incorporating adversarial training to enhance robustness. Moreover, we integrate interpretability into the training process. The notion behind interpretability modeling is that rather than moving in the worst-case direction, the noise is confined to additional examples from current embeddings. We perform experiments on four datasets that demonstrate the superiority of Adversarial Neural-Brane over state-of-the-art algorithms.</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

Interpretable adversarial neural pairwise ranking for academic network embedding

  • 期刊论文

  • 发表日期：

    2025年1月6日

查看全文
  • Agyemang Paul,
  • Zhefu Wu,
  • Boyu Chen,
  • Kai Luo,
  • Luping Fang

摘要

Bayesian personalized ranking (BPR) has gained prominence as an effective method for pairwise learning, particularly in personalized tasks such as recommendation systems. However, recent developments in adversarial machine learning (AML) have raised concerns about the vulnerability of advanced BPR techniques. While adversarial training has proven effective in enhancing resilience and performance in the recommendation system (RS) domain, its application in network embedding remains underexplored. This study addresses the gap by introducing the adversarial neural Bayesian ranking-based academic network embedding (Adversarial Neural-Brane). The proposed approach integrates a neural network model with BPR to generate latent representations of vertices, incorporating adversarial training to enhance robustness. Moreover, we integrate interpretability into the training process. The notion behind interpretability modeling is that rather than moving in the worst-case direction, the noise is confined to additional examples from current embeddings. We perform experiments on four datasets that demonstrate the superiority of Adversarial Neural-Brane over state-of-the-art algorithms.