Multi-label classification (MLC) is a common task across diverse real-world applications, including document tagging, disease gene prediction, and multimedia content annotation. However, the traditional MLC approaches struggle to balance classification accuracy with computational efficiency, especially when handling high-dimensional and intrinsically complex data. To resolve this, we propose a multi-label classification model, namely Multi-Label Granular Ball Twin Support Vector Machine (MLGBTSVM) by combining granular ball computing with multi-label twin support vector machine. Our model improves classification accuracy by combining granular ball abstraction with Twin Support Vector Machine’s (TSVM) dual non-parallel partitioning hyperplanes to effectively model multi-label data. It uses the Structural Risk Minimization principle, incorporating regularization terms, which effectively mitigate the problem of overfitting. MLGBTSVM also constructs classifiers based on the coarse-to-fine structure of granular balls, thereby reducing dependence on specific data instances and efficiently addressing the optimization problem. A comprehensive evaluation on benchmark datasets reveals that the proposed algorithm exhibits superior generalization performance compared to existing baseline methods.