A novel method of unsupervised feature selection UFS-IR based on improved ReliefF is proposed to solve the problem of lack of category information in feature selection. As the ReliefF algorithm has a small sampling probability of small class samples, it cannot delete the defects of redundant features. This method uses DBSCAN clustering algorithm to guide the classification. By improving the sampling strategy, it uses the adjusted cosine similarity to measure the correlation between features as a de-redundancy credential. Experiments show that UFS-IR can effectively reduce the data dimension while ensuring the maximum correlation redundancy of the feature subset, and with good performance.