In this paper, we consider the image classification problem. Unlike conventional local learning technique, a novel framework, which is based on the proposed sparsity induced neighbors (SINs) instead of widely used k nearest neighbors, is presented. Within this framework, the SINs of test image are training images associated with the nonzero entries in the sparse representation of test image, and they can be found by using kernel sparse coding algorithm. While its SINs are weighted properly, the test image can be classified as the category that is assigned the most weights. Moreover, we also apply the label embeddings learning in the framework, to model the similarity between categories and improve discriminative performance. Experimental results show that the proposed method can achieve state-of-the-art performance on three commonly-used datasets.