Abstract: To improve the accuracy and stability of feature selection in label-free scenarios, an unsupervised feature selection algorithm based on feature clustering and isometric mapping was proposed. Feature clustering clustered features with high similarity into one class, and a new feature score measurement function was defined by combining isometric mapping and sparse coefficient matrix. This function scored the features in each feature cluster and selected the representative features with the highest scores in each class cluster to form a feature subset. Experimental results on fourteen widely used datasets show that the proposed algorithm can select features with strong classification ability and the algorithm is highly generalizable.