Advances in meta-learning and zero-shot learning for multi-label classification: A review

Abstract

Effectively dealing with multi-label classification is a significant challenge. Traditional methods often struggle with issues such as label dependencies, data imbalance, and a limited number of annotated datasets. However, meta-learning and zero-shot learning models offer promising solutions by leveraging previous tasks to enable rapid generalization with minimal data. In this paper, we provide a comprehensive review of meta-learning and zero-shot strategies for multi-label classification in various domains, including audio, text, image, and sensor data, focusing on research published between 2019 and 2025. It presents an overview of commonly used datasets and a detailed description of models designed to capture the relationships inherent in multi-label scenarios. In addition, we propose a novel categorization framework based on neural architecture enhancements, algorithm adaptation, and problem transformation to highlight the main contributions of the reviewed literature. The aim of this review is to provide valuable insights into the current state of meta-learning and zero-shot approaches for multi-label classification, offering guidance for future research and development in addressing the complexities of real-world multi-label tasks.