Scholarlink

<p>This paper addresses a key challenge in multi-source transfer learning, where raw data from certain sources may be inaccessible due to privacy, security, or storage constraints, leaving only pre-trained models available. To overcome this limitation, we propose a novel framework called LNF -MSTL (Limited-Access Fuzzy Multi-Source Transfer Learning), designed for regression problems under restricted data access. Unlike existing methods that assume complete access to all source data or directly reuse pre-trained models, LAF-MSTL introduces a two-level adaptation mechanism: (i) latent-space alignment and clustering-guided fuzzy rule selection for sources with full data access, and (ii) pseudo-label refinement and fuzzy rule integration to extract and adapt knowledge from pre-trained models. By integrating neural networks, fuzzy clustering, and Takagi–Sugeno (TS) fuzzy rules, the framework effectively manages uncertainty and facilitates systematic knowledge transfer across heterogeneous domains. Experiments on four real-world regression datasets demonstrate that LAF-MSTL consistently outperforms state-of-the-art multi-source transfer learning techniques in limited-access scenarios. These results confirm the framework’s capability to achieve accurate and reliable knowledge transfer when multi-source information is incomplete.</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

Multi-source transfer learning with limited data access: a neural networks and fuzzy rules approach for regression problems

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Masoume Gholizade,
  • Hadi Soltanizadeh,
  • Mohammad Rahmanimanesh

摘要

This paper addresses a key challenge in multi-source transfer learning, where raw data from certain sources may be inaccessible due to privacy, security, or storage constraints, leaving only pre-trained models available. To overcome this limitation, we propose a novel framework called LNF -MSTL (Limited-Access Fuzzy Multi-Source Transfer Learning), designed for regression problems under restricted data access. Unlike existing methods that assume complete access to all source data or directly reuse pre-trained models, LAF-MSTL introduces a two-level adaptation mechanism: (i) latent-space alignment and clustering-guided fuzzy rule selection for sources with full data access, and (ii) pseudo-label refinement and fuzzy rule integration to extract and adapt knowledge from pre-trained models. By integrating neural networks, fuzzy clustering, and Takagi–Sugeno (TS) fuzzy rules, the framework effectively manages uncertainty and facilitates systematic knowledge transfer across heterogeneous domains. Experiments on four real-world regression datasets demonstrate that LAF-MSTL consistently outperforms state-of-the-art multi-source transfer learning techniques in limited-access scenarios. These results confirm the framework’s capability to achieve accurate and reliable knowledge transfer when multi-source information is incomplete.