Journal of Computer Applications

    Next Articles

Combining preprocessing methods and adversarial learning for fair link prediction

  

  • Received:2024-08-08 Revised:2024-11-04 Online:2024-11-19 Published:2024-11-19
  • Contact: ZHU Yan
  • Supported by:
    Sichuan Provincial Science and Technology Plan Project

结合预处理方法和对抗学习的公平链接预测

彭一峰1,朱焱2   

  1. 1. 西南交通大学
    2. 西南交通大学 信息科学与技术学院,成都 610031;
  • 通讯作者: 朱焱
  • 基金资助:
    四川省科技项目

Abstract: Abstract:Link prediction is a crucial task in network analysis that explores interactions between entities and forecasts potential new relationships in evolving networks. Link prediction may exhibit biases, especially concerning links between entities with sensitive attributes. For example,the issue of "filter bubbles", which amplify the isolation of publicly accessible information online and reduce diversity. To address these challenges, The ALFLP (Adding Link and Adversarial Learning for Fair Link Prediction) method was proposed, which combines preprocessing techniques to balance link density among disadvantaged link groups and adversarial learning methods during processing. The adversarial learning approach involves a generator-discriminator interplay to facilitate more inter-group links, thereby alleviating filter bubble effects. The experimental results on real datasets (pokec_n, pokec_z) show that, compared with baseline methods such as Jaccard, the AUC index of the ALFLP method is improved by about 12 percentage points and 10 percentage points respectively on pokec_n and pokec_z, and the modred index is improved by about 0.14 and 0.1 respectively. The ALFLP method can achieve a better balance between fairness and prediction accuracy.

Key words: Keywords: link prediction, fairness, filter bubbles, preprocessing, adversarial learning

摘要: 摘 要: 链接预测是网络分析的一项重要任务,它研究个体之间的相互作用,并推断在未来不断发展的网络中可能出现的新关系。但链接预测的过程中可能存在偏见,特别是当涉及到个体之间的链接,且这些个体包含某些敏感属性时。如“过滤气泡”问题,这一效应体现在对在线用户公开信息的隔离程度加强和多样性减小。针对上述问题,从算法公平性的角度缓解“过滤气泡”问题,将预处理阶段的方法和处理阶段的方法相结合,提出ALFLP(Adding Link and Adversarial Learning for Fair Link Prediction)方法。在预处理阶段,通过向劣势链接群组添加链接,减少不同群组链接密度差异;在处理阶段,将预处理阶段的输出输入到基于对抗学习的方法中,通过生成器与鉴别器相互博弈,促进更多组间链接,从而缓解“过滤气泡”问题。在真实数据集(pokec_n、pokec_z)上进行的实验结果表明,与其他基线方法(如Jaccard)相比,ALFLP方法的AUC指标分别提高了约12个百分点和10个百分点,modred指标分别提高了约0.14和0.1。ALFLP方法能在公平性和预测准确性之间实现较好的权衡。

关键词: 关键词: 链接预测, 公平性, 过滤气泡, 预处理, 对抗学习

CLC Number: