Abstract: In the area of feature selection, most existing methods cannot simultaneously capture the importance of different features and enforce orthogonal constraint on the projected subspace to improve the discrimination of selected features. To deal with this issue, a new unsupervised feature selection method called joint graph embedding and feature weighting (JGEFW) is proposed. To be specific, graph-embedding local structure learning is leveraged to obtain the similarity matrix and the cluster indicator matrix. Moreover, the weight matrix is learned through the orthogonal regression, to select the discriminative and non-redundant features. At last, an alternating iteration algorithm is developed for solving the proposed objective function. Finally, experimental results about clustering evaluation metrics on four publicly available datasets demonstrate that JGEFW is better than others in most cases.

Key words: feature selection, feature weight, unsupervised learning, graph embedding