Feature selection not only enhances the efficiency of regression modelling but also reduces the detrimental effects of feature redundancy and noises. This paper proposes a multi-target feature selection algorithm based on adaptive graph learning. Specifically, the method imposes a low-rank constraint on the regression matrix, enabling simultaneous modelling of inter-target, input-output and inter-sample relationships within a general framework. The similarity-induced graph matrix is learned to adaptively preserve samples' similarity structure to alleviate the influence of noises and outliers. Furthermore, we introduce a manifold regularizer to preserve the global target correlations to ensure the global target correlations structure of data in the subsequent learning process. An alternative optimization algorithm is presented to solve the final objective function. Extensive experiments conducted on real-world data sets demonstrate that the proposed method is superior to state-of-the-art multi-target feature selection methods.