Unsupervised feature selection based on dual-graph clustering learning and adaptive weighting

In unsupervised feature selection, existing algorithms tend to prioritize the manifold structure inherent in the data space, yet overlook the manifold-related information within the feature space. This results in incomplete excavation of the intrinsic data structure, thereby affecting the accuracy of feature selection. To fully leverage the dual-space manifold structures of data and features, this paper proposes an unsupervised feature selection algorithm based on dual-graph clustering learning and adaptive weighting (UDLA). Firstly, UDLA retains the manifold structure of data in both the data space and the feature space, and integrates the dual-structure manifold information into the submatrices of clustering pseudo-labels. The implementation of dual-graph clustering learning facilitates a more thorough utilization of the manifold information derived from both the data space and the feature space. Secondly, UDLA performs weighting processing on the original space and the subspace generated by non-negative matrix factorization. It treats the decomposed submatrices as the feature transformation matrix and carries out manifold learning on this matrix. Through adaptive weighting learning, UDLA can select features with higher discriminative power from the original data space, and more critical feature subsets can be identified simultaneously. Finally, by combining the l_2,1−2-norm with minimum redundancy to constrain the feature transformation matrix, UDLA can select features that exhibit strong relevance. As a result, a feature subset characterized by lower sparsity and reduced redundancy is achieved.