[Paper Review] A Survey on Negative Transfer

Wen Zhang, Lingfei Deng+2 more|arXiv (Cornell University)|Sep 2, 2020

Domain Adaptation and Few-Shot LearningComputer Science118 references19 citations

TL;DR

This paper presents the first comprehensive survey on negative transfer (NT) in transfer learning, systematically reviewing over 50 approaches across four categories: secure transfer, domain similarity estimation, distant transfer, and NT mitigation. It identifies key factors causing NT and provides guidelines for selecting appropriate strategies based on domain similarity, significantly advancing the understanding and practical handling of NT in real-world transfer learning applications.

ABSTRACT

Transfer learning (TL) utilizes data or knowledge from one or more source domains to facilitate the learning in a target domain. It is particularly useful when the target domain has very few or no labeled data, due to annotation expense, privacy concerns, etc. Unfortunately, the effectiveness of TL is not always guaranteed. Negative transfer (NT), i.e., leveraging source domain data/knowledge undesirably reduces the learning performance in the target domain, has been a long-standing and challenging problem in TL. Various approaches have been proposed in the literature to handle it. However, there does not exist a systematic survey on the formulation of NT, the factors leading to NT, and the algorithms that mitigate NT. This paper fills this gap, by first introducing the definition of NT and its factors, then reviewing about fifty representative approaches for overcoming NT, according to four categories: secure transfer, domain similarity estimation, distant transfer, and NT mitigation. NT in related fields, e.g., multi-task learning, lifelong learning, and adversarial attacks, are also discussed.

Motivation & Objective

• To address the long-standing challenge of negative transfer (NT) in transfer learning, where source domain knowledge harms target domain performance.
• To provide a systematic review of over 50 representative approaches to NT mitigation, filling a critical gap in existing literature.
• To categorize and analyze NT mitigation techniques based on four key strategies: secure transfer, domain similarity estimation, distant transfer, and NT mitigation.
• To offer practical guidelines for selecting transfer strategies based on estimated domain similarity.
• To identify open challenges and future research directions in NT, including regression problems, open-set scenarios, and theoretical diagnostics of NT factors.

Proposed method

• The paper classifies NT mitigation approaches into four categories: secure transfer, domain similarity estimation, distant transfer, and NT mitigation, based on their underlying mechanisms.
• It evaluates methods using a structured framework that considers transferability, domain divergence, and data quality factors.
• The authors analyze techniques such as feature adaptation, model adaptation, adversarial training, and instance-level reweighting to assess their effectiveness in reducing NT.
• A comparative analysis is conducted across 69 representative works, with performance evaluated on criteria including domain similarity, transfer algorithm, and data quality.
• The survey includes theoretical and empirical assessments of methods, focusing on those explicitly addressing NT from a principled perspective.
• Guidelines are derived from the analysis, recommending strategy selection based on estimated domain similarity (high, medium, low).

Experimental results

Research questions

• RQ1 What are the primary causes and contributing factors of negative transfer in transfer learning, particularly in classification tasks?
• RQ2 How do different categories of NT mitigation—secure transfer, domain similarity estimation, distant transfer, and NT mitigation—perform under varying domain similarity conditions?
• RQ3 Which transfer learning paradigms (e.g., unsupervised domain adaptation, few-shot learning) are most vulnerable to NT, and how can they be secured?
• RQ4 What are the key limitations of current NT mitigation methods, especially in regression and open-set learning scenarios?
• RQ5 How can the exact factors leading to NT be diagnosed in real-world applications to guide optimal strategy selection?

Key findings

• Most current NT research focuses on NT mitigation and domain similarity estimation, with a strong emphasis on data transferability enhancement.
• NT mitigation approaches predominantly target data transferability, often combining domain divergence, transfer algorithm, and data quality factors.
• Secure transfer strategies are largely based on model adaptation, suggesting a need for more diverse secure transfer mechanisms.
• The majority of NT mitigation methods consider multiple factors (e.g., domain divergence, data quality), indicating a trend toward holistic mitigation strategies.
• There is a notable lack of NT mitigation research for regression problems, highlighting a significant gap in current literature.
• The survey identifies that theoretical diagnostics of NT factors remain underdeveloped, suggesting a need for new empirical and theoretical tools to identify root causes in specific applications.