A Method for Extraction of Causal Relation between Kansei and Design Attributes Considering Visual Attention (An Approach Based on Gaze Feature and Rough Set Theory)

In the design of kansei (emotional) quality, one of the important issues is to extract causal relations between physical design attributes and the customer's emotional responses. Without such relations, a designer has to rely on his/her own sense that may be different from the customer's. In this paper, we propose a new method for extraction of logical rules consisting of combinations of design attributes that explain a customer's emotional judgment towards product appearance. In the method, we apply a reduct calculation in rough set theory to derive alternatives of causal rules between design attributes and emotional judgments, and use the customer's eye gaze features for refining the rules. We extract two types of visual attentions (VA), i.e., a single visual attention (SVA) and a combinational visual attention (CVA), by using the proposed gaze features. To demonstrate the effectiveness of the method, we conducted a sensory evaluation experiment using a car-interior design as a case study. In the experiment, multiple participants evaluated impressions of multiple design samples by selecting from a set of words. During the experiment, we recorded the participants' eye gaze movements as coordinates on a screen, and asked them to vocalize aloud what they were thinking. After an evaluation of each design sample, we conducted a retrospective interview. From the results, we confirmed that the estimated SVA and CVA significantly covered the vocalized thoughts and statements made in the retrospective interview. The estimated VA reduced 53% of the erroneous causal rules and improved the quality of the rules. We found a case where two participants making the same emotional judgments have implicitly different points of view when evaluating the same design sample. Most conventional causality analysis has been unsuccessful in finding such diversity of points of view.