A Survey of Mix-based Data Augmentation: Taxonomy, Methods, Applications, and Explainability

TL;DR

This survey provides the first systematic, taxonomy-driven review of MixDA, detailing Mixup-based, Cutmix-based, and other mixing approaches, their theoretical underpinnings in vicinal risk minimization and regularization, and their broad applicability across vision, language, graphs, and beyond. It consolidates three dozen core techniques, analyzes their mechanisms for improving generalization and calibration, and discusses practical considerations such as sample selection, saliency guidance, and diversity. The work also surveys a wide range of applications—semi-supervised learning, contrastive learning, adversarial training, domain adaptation, and more—and offers critical insights into explainability and the remaining challenges. Altogether, it highlights MixDA’s versatility, efficiency, and potential as a unifying framework for robust, data-efficient learning, while outlining concrete directions for future research, including test-time mixing, integration with large models, and modality-specific adaptations.

Abstract

Data augmentation (DA) is indispensable in modern machine learning and deep neural networks. The basic idea of DA is to construct new training data to improve the model's generalization by adding slightly disturbed versions of existing data or synthesizing new data. This survey comprehensively reviews a crucial subset of DA techniques, namely Mix-based Data Augmentation (MixDA), which generates novel samples by combining multiple examples. In contrast to traditional DA approaches that operate on single samples or entire datasets, MixDA stands out due to its effectiveness, simplicity, flexibility, computational efficiency, theoretical foundation, and broad applicability. We begin by introducing a novel taxonomy that categorizes MixDA into Mixup-based, Cutmix-based, and mixture approaches based on a hierarchical perspective of the data mixing operation. Subsequently, we provide an in-depth review of various MixDA techniques, focusing on their underlying motivations. Owing to its versatility, MixDA has penetrated a wide range of applications, which we also thoroughly investigate in this survey. Moreover, we delve into the underlying mechanisms of MixDA's effectiveness by examining its impact on model generalization and calibration while providing insights into the model's behavior by analyzing the inherent properties of MixDA. Finally, we recapitulate the critical findings and fundamental challenges of current MixDA studies while outlining the potential directions for future works. Different from previous related surveys that focus on DA approaches in specific domains (e.g., CV and NLP) or only review a limited subset of MixDA studies, we are the first to provide a systematical survey of MixDA, covering its taxonomy, methodology, application, and explainability. Furthermore, we provide promising directions for researchers interested in this exciting area.

Figures (4)

• Figure 1: The sketch of Mixup. Each pixel in the Mixup-generated image is a convex combination of the corresponding pixels from two randomly sampled images with the mix ratio $\lambda$.
• Figure 2: An example of Cutmix with mix ratio $\lambda=0.25$, in which $1/4$ area of the dog image (upper right) is cut and pasted onto the corresponding location of the cat (upper right).
• Figure 3: Illustration of why Cutmix is problematic when the cut-and-paste patch contains no information about the dog. However, based on the label combination policy, the probability of the dog for the new image is non-zero ($0.25$), misleading the learned model.
• Figure 4: Illustration of SSMix. The $j$-th sentence's most irrelevant tokens are replaced with the $i$-th sentence's most related tokens.