A Soft subspace clustering algorithm based on multi-objective optimization and reliability measure

Abstract

Subspace clustering finds clusters in subspaces of the data instead of the entire data space to deal with high-dimensional data. Most existing subspace clustering algorithms lean on just one single objective function. Single objective function is often biased. On the other hand, most existing subspace clustering algorithms are based on wrapper approach, which brings a negative effect on the quality of subspace clustering. This paper presents a soft subspace clustering algorithm based on multi-objective evolutionary algorithm and reliability measure, called R-MOSSC. Comparing with optimization of a scalar function combining multiple objectives, it does not need to determine weight hyperparameters, and offers a deep insight into the problem by obtaining a set of solutions. Further, reliability-based dimension weight matrix from filter approach is used to enhance the performance of subspace clustering. Simulation results show that R-MOSSC is better than existing algorithms. © Springer-Verlag Berlin Heidelberg 2013.