Unsupervised Feature Selection (UFS) aims to eliminate redundant and noisy features from unlabeled data to improve data quality. However, most existing UFS methods rely on a single evaluation criterion to select features, such as preserving structural similarity, maintaining clustering structure, or minimizing reconstruction error, which limits their ability to fully exploit the rich and diverse information embedded in features. To overcome these limitations, this paper proposes Multi-Information Fusion Unsupervised Feature Selection (MIF-UFS). Specifically, the model incorporates both k-means clustering and graph manifold structure into a unified reconstruction framework. This design preserves the structural relationships in the original feature space while guiding the model to learn more discriminative clustering representations in the embedding space. Subsequently, the Alternating Direction Method of Multipliers (ADMM) is employed to efficiently optimize the proposed model. Experiments on several benchmark datasets demonstrate that our method outperforms existing state-of-the-art approaches.