Scholarlink

<p>Privacy-preserving collaborative filtering (PPCF) systems are effective tools for providing personalized recommendations to users without violating their privacy. However, existing data disguising methods, which use randomized perturbation techniques (RPT) to achieve high privacy levels, often corrupt real user data, resulting in significant losses in prediction accuracy. To address this issue, this study proposes an autoencoder(AE)-based privacy-preserving recommendation approach. The experimental works conducted on two real datasets show that the AE-based method provides significantly better accuracy than traditional neighborhood-based and matrix factorization techniques, especially at high privacy levels. We also conduct a comparative analysis of the prediction accuracy enhancements provided by the proposed AE model under diverse architectural settings, such as variations in activation functions and the number of hidden layers. The analysis shows that two encoder layered-AE model with tanh activation provides the most accurate results on both datasets. Overall, the findings show that the proposed AE architecture successfully reduces RPT-induced distortions and provides a strong alternative to traditional PPCF techniques.</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

Privacy-preserving autoencoder-based collaborative filtering

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Alper Yargıç,
  • Elif Tuğçe Açıl,
  • Zeynep Batmaz

摘要

Privacy-preserving collaborative filtering (PPCF) systems are effective tools for providing personalized recommendations to users without violating their privacy. However, existing data disguising methods, which use randomized perturbation techniques (RPT) to achieve high privacy levels, often corrupt real user data, resulting in significant losses in prediction accuracy. To address this issue, this study proposes an autoencoder(AE)-based privacy-preserving recommendation approach. The experimental works conducted on two real datasets show that the AE-based method provides significantly better accuracy than traditional neighborhood-based and matrix factorization techniques, especially at high privacy levels. We also conduct a comparative analysis of the prediction accuracy enhancements provided by the proposed AE model under diverse architectural settings, such as variations in activation functions and the number of hidden layers. The analysis shows that two encoder layered-AE model with tanh activation provides the most accurate results on both datasets. Overall, the findings show that the proposed AE architecture successfully reduces RPT-induced distortions and provides a strong alternative to traditional PPCF techniques.