The core challenge in multi-view feature selection lies in identifying discriminative features that capture both consensus and diversity information. Feature selection methods based on consensus learning have garnered significant attention due to their ability to reveal the common underlying structure across different views. To effectively capture both homogeneous and heterogeneous information across views, this paper proposes a novel multi-view unsupervised feature selection method guided by diversity and consensus structure, referred to as GDCSMuFS. The diversity structure learning across multiple views is crucial for preserving the heterogeneous information in the selected features. Building on this concept, the proposed method begins by incorporating both homogeneous and heterogeneous information. It first learns the multi-view cluster structure matrix through non-negative matrix factorization. Then, multi-view spectral analysis is employed to integrate the cluster structure matrices into a consensus pseudo-label matrix, enabling the extraction of consensus information at the label level. Subsequently, a feature selection matrix is constructed under the shared constraints of consensus labels and graph structures. Finally, sparse constraints are applied to the feature selection matrix, and the model is solved using an alternating iterative algorithm. Extensive experiments on benchmark datasets demonstrate the effectiveness of the proposed method, with comparative results showing its superiority over existing algorithms.