Recently, nuclear norm based matrix regression (NMR) for classification has been proposed to characterize the whole structure of the error image. However, NMR ignores both the label information and the group structure of training samples. This paper presents a novel yet effective coding scheme called locality-constrained group sparse coding regularized NMR (LGNMR) which not only overcomes these limitations but also utilizes the similarities between test samples and training samples. We adopt the inexact augmented lagrange multiplier (IALM) method to solve the proposed model efficiently. Experiments on both Extended Yale B database and AR database have shown that the proposed method outperforms the state-of-the-art regression based classification methods.