Scholarlink

<p>Federated recommendation (FR) offers users personalized recommendation services while safeguarding their privacy. Conventional FR considers all items that users interact with as items that users like. However, this assumption fails to capture genuine user preferences due to some <i>noisy samples</i>, in which users interact with items they do not like. Most research in FR disregards such noisy samples, consequently inducing local models to learn inaccurate user preferences. Since the global model is aggregated from local models, its performance is compromised, adversely affecting user experience. Furthermore, data heterogeneity, privacy protection, and communication burden requirements make denoising FR more complicated. In this paper, we propose a <Emphasis Type="BoldItalic">S</Emphasis><i>elf</i>-<Emphasis Type="BoldItalic">T</Emphasis><i>raining</i> <Emphasis Type="BoldItalic">D</Emphasis><i>ual-Network</i> (STDFed) for denoising FR. Specifically, each client’s global and local models inherently form a dual-network. On each client, STDFed considers samples with high predicted probabilities from the dual-network as clean samples to construct a self-training dataset, while the remaining are treated as unlabeled samples. The self-training dataset replaces the original dataset to support local training. In subsequent rounds, STDFed identifies the unlabeled samples with low predicted probabilities as noisy samples, which will be either discarded or labeled as 0 and then added to the self-training dataset. STDFed is model-agnostic and can be applied to most FR. Extensive experiments show that STDFed improves the performance of FR by an average of 6.96% across multiple datasets without increasing communication costs.</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

Self-training dual-network for denoising federated recommendation

  • 期刊论文

  • 发表日期：

    2025年3月22日

查看全文
  • Pingshan Liu,
  • Haoning He,
  • Guoxin Lu

摘要

Federated recommendation (FR) offers users personalized recommendation services while safeguarding their privacy. Conventional FR considers all items that users interact with as items that users like. However, this assumption fails to capture genuine user preferences due to some noisy samples, in which users interact with items they do not like. Most research in FR disregards such noisy samples, consequently inducing local models to learn inaccurate user preferences. Since the global model is aggregated from local models, its performance is compromised, adversely affecting user experience. Furthermore, data heterogeneity, privacy protection, and communication burden requirements make denoising FR more complicated. In this paper, we propose a Self-Training Dual-Network (STDFed) for denoising FR. Specifically, each client’s global and local models inherently form a dual-network. On each client, STDFed considers samples with high predicted probabilities from the dual-network as clean samples to construct a self-training dataset, while the remaining are treated as unlabeled samples. The self-training dataset replaces the original dataset to support local training. In subsequent rounds, STDFed identifies the unlabeled samples with low predicted probabilities as noisy samples, which will be either discarded or labeled as 0 and then added to the self-training dataset. STDFed is model-agnostic and can be applied to most FR. Extensive experiments show that STDFed improves the performance of FR by an average of 6.96% across multiple datasets without increasing communication costs.