Multi-label classification learning provides a multi-dimensional perspective for polysemic object, and becomes a new research hotspot in machine learning in recent years. In the big data environment, it is urgent to obtain a fast and efficient multi-label classification algorithm. Kernel extreme learning machine was applied to multi-label classification problem (ML-KELM) in this paper, so the iterative learning operations can be avoided. Meanwhile, a dynamic, self-adaptive threshold function was designed to solve the transformation from ML-KELM network’s real-value outputs to binary multi-label vector. ML-KELM has the least square optimal solution of ELM, and less parameters that needs adjustment, stable running, faster convergence speed and better generalization performance. Extensive multi-label classification experiments were conducted on data sets of different scale. Comparison results show that ML-KELM outperformance in large scale dataset with high dimension instance feature.