Graph construction, an open and challenging problem, is of great significance for multi-view unsupervised feature selection. So far, many graph construction methods, such as distance-based (the local structure) and self-reconstruction-based (the global structure), have been devised to serve the feature selection task. Although these methods have achieved some improvements, they fail to utilize high-order neighbor information, let alone exploit the neighbor information of different orders, to improve the feature selection task. In this paper, we propose a new insight to construct graphs that can accommodate multi-order neighbor information for selecting the relevant features. Besides, we observe that existing methods adopts the general information fusion strategy in multi-view learning, e.g. fusing graphs, without taking into account the unique characteristics of the feature selection task. Therefore, the proposed method seeks to project multi-view data onto a shared latent representation, which explores the complementarity tailored to the feature selection task at the feature level. A simple yet effective algorithm is designed to solve the optimization problem of the objective function. Extensive clustering experiments demonstrate the superiority of our method over state-of-the-art ones.