Graph-based semi-supervised classification has attracted extensive attention in the field of pattern recognition and machine learning in recent years. However, many traditional methods use fixed neighborhood size when constructing neighborhood graph. Besides, they treat all samples equally in the process of model training, and ignore the differences among samples, which affects the effectiveness of the methods. In this paper, we propose a semi-supervised classification method based on self-paced learning and sparse self-expression. The method extracts and maintains the discriminative sparse self-expression structure of data, and proposes a new self-paced function based on self-paced learning mechanism, which combines the soft weight with the hard weight of data importance. The proposed method can adaptively establish the relationship between data points, and automatically give the data importance, which guides the learning process of partially labeled training data sequentially from the simple to the complex. Besides, the proposed method has multi-class explicit nonlinear classification function. The experimental results on several benchmark datasets show that the proposed algorithm has satisfactory semi-supervised classification effects.