Feature Selection for Local Learning Based Clustering

For most clustering algorithms, their performance will strongly depend on the data representation. In this paper, we attempt to obtain better data representations through feature selection, particularly for the Local Learning based Clustering (LLC) [1]. We assign a weight to each feature, and incorporate it into the built-in regularization of LLC algorithm to take into account of the relevance of each feature for the clustering. Accordingly, the weights are estimated iteratively with the clustering. We show that the resulting weighted regularization with an additional constraint on the weights is equivalent to a known sparsepromoting penalty, thus the weights for irrelevant features can be driven towards zero. Experiments on several benchmark datasets demonstrate the effectiveness of the proposed method.