Abstract In this paper, we propose an effective approach to semi-supervised classification through kernel-based sparse representation. The new method computes the sparse representation of data in the feature space, and then the learner is subject to a cost function which aims to preserve the sparse representing coefficients. By mapping the data into the feature space, the so-called “l2-norm problem” that may be encountered when directly applying sparse representations to non-image data classification tasks will be naturally alleviated, and meanwhile, the label of a data point can be reconstructed more precisely by the labels of other data points using the sparse representing coefficients. Inherited from sparse representation, our method can adaptively establish the relationship between data points, and has high discriminative ability. Furthermore, the new method has a natural multi-class explicit expression for new samples. Experimental results on several benchmark data sets are provided to show the effectiveness of our method.