Adaptive Feature-Weighted Local-Global Clustering

Clustering has long been a fundamental problem in machine learning and data mining, with the aim of grouping data samples on the basis of their intrinsic similarity. However, the consensus that different features often exhibit varying levels of discriminative power in clustering model learning is under explored sufficiently in collaboration with the pseudo-label guided unsupervised discriminative analysis. To this end, we propose an Adaptive Feature-Weighted Local-global data Clustering (AFW-LGC) model which is featured by two improvements. First, AFW-LGC takes into account both global separability (between-cluster scatter) and local compactness (within-cluster scatter) whose impacts are mediated by a learnable parameter. Second, the different contributions of features are adaptively learned in AFW-LGC for further discriminative ability enhancement. Both improvements are seamlessly integrated for feature-weighted unsupervised discriminative subspace clustering nature of AFW-LGC. Extensive experiments on eight data sets demonstrate the superior clustering performance of AFW-LGC over some SOTA methods as well as the rationality of our proposed feature importance exploration strategy.