Publication

Abstract : High-dimensional data suffers from the curse of dimensionality and sparsity problems. Since all samples seem equidistant from each other in high-dimensional space, low-dimensional structures need to be found for cluster formation. This paper proposes a top-down approach for subspace clustering called projective clustering to identify clusters in low-dimensional subspaces using best low-rank matrix factorization strategy, singular value decomposition. The advantages of this approach are twofold. First is to obtain multiple low-dimensional substructures using the best low-rank approximation, thereby reducing the storage requirements. Second is the usage of the obtained projective clusters to retrieve approximate results of a given query in time-efficient manner. Experimentation on six real-world datasets proves the feasibility of our model for approximate information retrieval.