Unsupervised multi-view feature selection has become an important research direction in the field of pattern recognition and machine learning. However, most of existing methods fail to consider the redundancy information within and between views or the noise information in each view. In this paper, we propose a multi-level regularization-based unsupervised multi-view feature selection with adaptive graph learning. Our method adaptively learns a proper similarity matrix in the reduced feature space based on a learned projection matrix. To reduce the redundancy and noise information in the multi-view data, we adopt a multi-level regularization, which explores the structural sparsity, dependency, diversity information of the multi-view data, to constrain the learned projection matrix. Based on the obtained projection matrix, we rank the features and perform multi-view feature selection. We develop an effective iteration optimization algorithm to solve our method. A large number of experiments conducted on six popular multi-view datasets show that our method obtains excellent clustering performance and has superiority in comparison with mainstream methods.