Multi-Source Transfer Learning Based on Pre-Trained Models: A Fuzzy Rule-Based Approach

Gholizadeh, Masoome; Soltanizadeh, Hadi; RahmaniManesh, Mohammad

Signal and Data Processing Journal A scientific journal officially licensed by the Commission for Scientific Publications of the (MSRT). Publisher: Research Ceter for Developmen of Technologies

EN FA

Volume 23, Issue 1 (6-2026) JSDP 2026, 23(1): 83-0 | Back to browse issues page

Mendeley

Zotero

RefWorks

Gholizadeh M, Soltanizadeh H, RahmaniManesh M. Multi-Source Transfer Learning Based on Pre-Trained Models: A Fuzzy Rule-Based Approach. JSDP 2026; 23 (1) : 6
URL: http://jsdp.rcisp.ac.ir/article-1-1471-en.html

Multi-Source Transfer Learning Based on Pre-Trained Models: A Fuzzy Rule-Based Approach

Masoome Gholizadeh

, Hadi Soltanizadeh ^*

, Mohammad RahmaniManesh

Semnan Univeristy

Abstract: (14 Views)

In recent years, multi-source transfer learning (MSTL) has emerged as a promising paradigm for leveraging knowledge from multiple domains to improve learning performance in a target domain. However, a critical challenge arises when access to raw data from some source domains is restricted or entirely unavailable—a situation common in real-world applications due to privacy regulations, data ownership constraints, storage limitations, or security concerns. In such scenarios, only pre-trained models from certain sources may be accessible, significantly complicating the transfer of knowledge and amplifying the uncertainty involved in downstream tasks. To address these limitations, this paper introduces a novel MSTL framework that synergistically integrates neural networks, fuzzy clustering techniques, and Takagi-Sugeno fuzzy inference systems. The proposed method is designed to enhance the robustness of knowledge transfer under partial observability conditions while effectively managing the inherent uncertainty that arises from incomplete source information. Fuzzy rules are leveraged to model complex and nonlinear relationships among data features and to enable interpretability and flexibility in the knowledge adaptation process. Through fuzzy clustering, latent structures within the available source model outputs are identified, supporting more accurate alignment between the source and target domains. Our method is evaluated on four benchmark regression datasets, where the target domain suffers from limited or uncertain source knowledge. Experimental results demonstrate that the proposed approach consistently outperforms several state-of-the-art MSTL methods, particularly in cases where the amount of accessible source data is insufficient. This superior performance highlights the effectiveness of combining fuzzy logic with neural models for uncertainty-aware transfer learning. Moreover, the approach shows strong generalization capabilities and practical potential for deployment in sensitive and data-constrained environments, such as healthcare, finance, and cybersecurity.

Article number: 6

Keywords: Multi-source transfer learning, fuzzy rules, domain adaptation, pre-trained models, Takagi-Sugeno fuzzy rules.

Type of Study: Applicable | Subject: Paper
Received: 2025/05/31 | Accepted: 2026/02/8 | Published: 2026/09/16 | ePublished: 2026/09/16

Send email to the article author

Rights and permissions
	This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.