Reduction of Attribute Values for Kansei Representation

Abstract

In this paper, we discuss attribute-value reduction for Kansei analysis using information granulation. In the traditional "reduction" sense, the goal is to find the smallest number of attributes enabling us to discern each tuple or each decision class. Once we focus on the number of attribute values, i.e., the size and/or resolution of each attribute domain, another criterion is needed. We must ask ourselves which is better: 1) discerning them with a single attribute described in detail, or 2) discerning them with a few attributes described roughly. Our study answers this question. If we evaluate this difference in the light of understandability or of Kansei representation, we may prefer working with a few attributes described roughly because they yield simpler descriptions. To do this, we propose a criterion and an algorithm to find near-optimal solutions for the criterion, detailing results for databases in the UCI Machine Learning Repository.