PLS-based approach for Kansei analysis

A residential garden contributes to mental health in modern life. Gardening is a common recreational activity. From the view of the Kansei engineering, designing the garden is a quite difficult subject. Since garden components such as stones and trees are widely diversified, then number of possible design elements becomes quite large. Meanwhile, evaluation samples that can be used for Kansei Evaluation are limited. Relations between Kansei word evaluation and design elements had been analyzed with Quantification Theory type I, which is a variation of a multiple regression model. Since QT1 is based on the least square method, number of evaluation samples should be larger than the number of design elements. Thus, QT1 is not applicable in this case. Recently, PLS (Partial Least Squares) is becoming popular in the field of Chemometrics, which deal with extremely large number and interacted predictor variables. In this study, we utilized PLS for analyzing Kansei evaluation on residential gardens and their 89 design elements. Analyzing results of PLS and QT1 are compared. QT1 analyses were done on 5-fold design elements. Even when incorporating 89 variables, PLS’s multiple correlation coefficient was much higher than QT1. Analyzing result was made into hand-made virtual reality Kansei engineering system. The system contains two projectors and a PC. 3D models of parts such as trees and stones are dynamically chosen and allocated in the scene. The system was based on originally developed 3D computation and rendering library on Java. I. GENERAL INTRODUCTIONS A residential garden is the small personal garden in the yard of a residence. At least it was common in Romans and spread in the Renaissance era. In Japan, private and temple garden has more than 1000 years tradition. In medieval age, art of tea became common culture for Bushi, Japanese warrior, then small (often courtyard) tea garden spread to Bushi residence. In modern age of Japan from Meiji (1868-1912), Taisho (1912-1926), residential garden was popular for richer people, then gradually spread to common peoples houses [1]. In this modern world, people expect mellower, peace of mind, even curative atmosphere to the residential garden. Revealing relations between mellow and curative Kansei to garden design elements helps garden designing which accustomed to a client. Realization of the design with 3D graphics will contribute to make ideal garden design. It will stimulate interaction between a client and a garden designer. The steps of this research are as follows; • Survey of residential gardens • Analysis of Kansei evaluation data of residential gardens • Making a virtual reality system that assists realization of design from Kansei Modern society has complex stress and pressures. Most of the people require stress management. Turning entire living environment into a curative environment contributes to design of healthy urban environment. To study the garden in Kansei engineering methodology, there are several challenges. One is a large number of design elements. In this case, number of design elements is 32 items 89 categories. Although we have used 47 gardens for evaluation sample in Kansei evaluation experiment, the sample number is not sufficient for regression analysis which based on least-square method. We utilized PLS based analysis for this project. Another challenge is building a virtual reality Kansei system. Available VR systems require large amount of budget, space and efforts for model building. To encourage Kansei VR system for gardening, the VR system should be lowbudget and should base on common devices. In addition, model building has to be easier. Then, we have decided to make hand made VR system that can read and display 3D objects those are expressed in a common 3D format, OBJ. Relations between Kansei word evaluation and design elements had been analyzed with Quantification Theory type I, which is created by Japanese statistician Chikio Hayashi in 1950s. Evaluation values on a Kansei word are assigned to y (dependent) variable and design elements are assigned to x (independent) variables with dummy variables. In QT1, the qualitative variable such as product body color is called “ item”. The variations of the item such like white / blue / red are called“categories”. Categories are expressed with dummy variables those have 1/0 value. Weights to the categories will be found with a multiple regression model. Computing method is solving (number of all categories 1) simultaneous equations [2]. QT1 is a deterministic method because it is a variation of multiple regression model and its solving method uses least square method. Although QT1 is widely used, there are two defects: 1. Problem of sample size; QT1 incorporates (number of category 1) dummy variables. In a multiple regression model, simultaneous equations could not solved when number of variables exceed to the number of samples. Kansei engineering situation has tens of samples, and many cases Fifth International Workshop on Computational Intelligence & Applications IEEE SMC Hiroshima Chapter, Hiroshima University, Japan, November 10, 11 & 12, 2009