Scholarlink

<p>Graph neural networks (GNNs) have been widely used in link prediction tasks due to their strong capability. However, they are susceptible to higher latency caused by redundant correlation of neighboring nodes in the graph, which poses a challenge for practical application. In contrast, multi-layer perceptron (MLP) layers provide a more efficient inference process with greater flexibility in resource-constrained environments, but also reduce some prediction accuracy. To combine the benefits of both, we propose a knowledge distillation framework consisting of teacher (GNNs-based) and student (MLP-based) models for better link prediction. Specifically, we first design an Adversarial Generative Discriminator (AGD) that trains the discriminators and generators in an adversarial manner to adaptively assist student model learning in various complex scenarios. Then, we propose the Feature Aggregation Module (FAM) to enrich the neighborhood information of the nodes and further enhance the acquired feature by the student model before the distillation starts. The results of the extended experiment show that our method exhibits superior performance in link prediction tasks, outperforming the baseline Linkless Prediction (LLP) model in both accuracy and efficiency, and achieving competitive inference efficiency in eight datasets under two different configurations.</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

Link prediction via adversarial knowledge distillation and feature aggregation

  • 期刊论文

  • 发表日期：

    2025年3月31日

查看全文
  • Wen Li,
  • Xiaoning Song,
  • Wenjie Zhang,
  • Yang Hua,
  • Xiaojun Wu

摘要

Graph neural networks (GNNs) have been widely used in link prediction tasks due to their strong capability. However, they are susceptible to higher latency caused by redundant correlation of neighboring nodes in the graph, which poses a challenge for practical application. In contrast, multi-layer perceptron (MLP) layers provide a more efficient inference process with greater flexibility in resource-constrained environments, but also reduce some prediction accuracy. To combine the benefits of both, we propose a knowledge distillation framework consisting of teacher (GNNs-based) and student (MLP-based) models for better link prediction. Specifically, we first design an Adversarial Generative Discriminator (AGD) that trains the discriminators and generators in an adversarial manner to adaptively assist student model learning in various complex scenarios. Then, we propose the Feature Aggregation Module (FAM) to enrich the neighborhood information of the nodes and further enhance the acquired feature by the student model before the distillation starts. The results of the extended experiment show that our method exhibits superior performance in link prediction tasks, outperforming the baseline Linkless Prediction (LLP) model in both accuracy and efficiency, and achieving competitive inference efficiency in eight datasets under two different configurations.