A neighborhood rough set model for attribute reduction without distance metric

Abstract

Neighborhood rough set (NRS) is an important extension of rough set theory, which can process continuous data directly without any prior knowledge. As far as we know, all present neighborhood rough set models are defined on distance metric (usually Euclidean distance), which makes neighborhood rough set model invalid in high-dimensional space due to "Curse of Dimensionality". Even in low-dimensional space, the performance of this model will be degraded due to the neglect of attribute weight by distance metric. This paper proposes a novel neighborhood rough set model based on space partition, and designs an attribute reduction algorithm based on this model. Experimental results on UCI benchmark datasets show that our algorithm performs better than the state-of-the-art NRS algorithms.