A0768
Title: Covariate-robust clustering
Authors: Yunjin Choi - University of Seoul (Korea, South) [presenting]
Abstract: Common clustering methods can be overly influenced by single covariates, leading to inconsistent results across different covariates if each covariate is analyzed separately. To address this issue, a novel clustering approach identifying clusters that are not dominated by a single covariate is proposed but rather captures the underlying structure across all covariates. This is achieved by minimizing the maximum error across covariates. In this approach, a covariate's error is the number of points assigned to clusters that deviate from their closest covariate-specific centers. The motivation is the customer segmentation based on reviews that consist of positive and negative comments. Existing methods, although employed on the whole comments, often yield clusters dominated by positive or negative comments. The approach addresses this by providing clustering that bridges the gap between results based solely on positive or negative reviews.
