Robust Neighborhood Covering Reduction with Determinantal Point Process sampling

Abstract

Neighborhood Covering Reduction (NCR) methods are too rigid to filter out redundant neighborhoods and sensitive to noise. To tackle this problem, we propose a flexible neighborhood reduction method based on Determinantal Point Process (DPP) sampling in this paper. Sampling from a DPP, the probabilities of subset selection are computed from the correlation matrix of items and the subsets of diverse items will be assigned high probabilities. We model the process of neighborhood selection as a DPP and thereby implement a NCR algorithm with DPP neighborhood sampling (NCRDPP). NCRDPP selects significant neighborhoods of both high quality and diversity to form concise coverings as the approximations of data distributions, which facilitate the model generalization and robustness of neighborhood-based learning. Experimental results verify the superiority of DPP sampling for neighborhood selection and the robustness of NCRDPP method for noisy data classification.