Free Essay

Hagft

In:

Submitted By biswarup82
Words 4374
Pages 18
A STUDY AFFECTIVE FACTOR EXTRACTION USING
METHODOLOGY FOR KANSEI ENGINEERING
Tsung-Hsing Wang
Department of Life Product Design,
SHU-TE University of Science and Technology, Taiwan ROC
Email: ch8136@stu. edu. tw
Abstract
In recent years, the study affective factor selection (AFs) for Kansei engineering
(KE) has been an important issue in the industrial design field. Consumers’ affective responses (CARs) are usually presented in the form of a choice of adjectives. Based on the KE concept, this study conducted Factor Analysis (FA), Clustering Analysis
(CA) and Procrustes Analysis (PA) to select the CARs from mobile phones product’s shape. First, in the initial stage of the study, 60samples of mobile phones were collected from the fashion market place. Twenty-two pairs of adjectives describing the mobile phones were used for a Semantic Differential (SD) experiment. K-means was implemented to find the clustering segmentations of the CARs according to the factor loading from FA, and to obtain representative pairs of adjectives within the clustering segmentations. In the meanwhile, PA was also used to decide adjective priorities according to the sorting rule. Finally, these two methods were analyzed and compared.
Keywords: Kansei engineering, Affective factor, Methodology, Clustering analysis,
Procrustes analysis.
Introduction
In the field of consumer market, the appearance of a product tends to be an important factor affecting

consumers’ purchasing decision making. If product designers can notice product forms features selection (PFFs), they can effectively meet the expectations of consumers. Therefore, during the development of a new product, it is

The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

206

an important issue to effectively meet

adjectives. The basic assumption of KE

consumers’ affective responses (CARs).
CARs are usually presented by adjectives. As a result, this study focused on the investigation of Affec-

is the existence of a cause-and-effect relationship between PFFs and CARs
(Han & Hong, 2003), as shown in
Figure 1.

tive Factor Selection (AFs). The investigation on product design should focus on consumers’ subjective feelings

Shinya
Nagasawa
(2004) summarized many studies concerning

and psychological needs. In recent years, KE-related theories and models are frequently used in the studies concerning development of new prod-

KE. Although many scholars published their studies, the observation showed that there are still some misunderstandings and problems in the

ucts. KE is a consumer-oriented technique transforming individuals’ feeling and image into product design. It

studies of KE in the field of product design, and the main cause is that there are still some controversies over the

assesses the adjectives used by consumers to describe the product modeling to understand the correlation between product modeling and

usefulness of KE. Owing to these controversies, more experts and scholars have discussed and studied KE more comprehensively.

Figure 1.

KES concept model (modified from Han & Hong, 2003).

There are only few KE-related

ponents of chairs and psychological evaluation variables of users where adjectives were used as input factors

studies in AFs. In relevant AFs studies,
Semantic differential (SD) and other statistical analysis methods were firstly used to summarize the variables of

and 3D image as output. Han (2003) verified that there is a cause-and-effect relationship between PFFs and CARs.
To obtain CARs, Osgood (1957) used

adjectives. Jindo (1995) treated office chairs as an example to investigate the relationship between appearance com-

SD to conduct an experiment where consumers were invited to select adjectives to assess product samples.

Literature Review

The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

207

Moreover, Salvador (2005) also used

technique to factor loadings can screen

SD to study cutting machines. Hsu
(2000) used SD to test the relationship between the appearance characteristics of telephone samples and psychological

out important clusters.

evaluation of users. Furthermore, in a sneakers-related study, Alca´ntara
(2005) used FA to define semantic axes

as well as to avoid the diversified structural data. The study using PA to select subsets of variables was firstly

and used SD to construct semantic meanings. Consumers’ perception and pref-

conducted by Krzanowski (1987). He also used Principal Component Analysis (PCA) to process some analytical data of Kansei. Andrade (2004) pro-

erence can represent the semantic meanings of products and can also strongly reflect their level of ac-

posed the “Procrustes rotation” and general appearance method to process two research cases. The concept of

ceptance of products. Hsu (2000) used
FA to investigate the difference between users and designers in CARs.

“Procrustes rotation” is a simple process and is extremely useful in the application of some Stoichiometry software. He also described the basic

In addition to FA, CA is also frequently used to analyze the data obtained from SD experiment. For example,
Chen (2010) used a new approach with

knowledge and some cases of the use of “Procrustes rotation” to demonstrate its actual application.

both breadth and priority. It is a graphic structure-based clustering method. In the use of FA, when the dimensions of variables are similar factors and the factor loadings are close, a satisfactory result cannot be obtained. Therefore,

Research process and analysis

CA is further used to analyze the factor loading obtained from the use of FA to obtain a more satisfactory result.
Vigneau (2003) suggested that the dimensions of variables can be arranged in to a homogenous group according to their factor loadings.
Moreover, Sahmer (2008) indicated that the application of clustering

The other approach is to use PA to select the key dimensions of variables,

Factor analysis and results
Factor analysis is a statistical method, which can reduce the number of variables and be used for categorization. It is frequently used in psychological studies in early days. It is now comprehensively applied to the studies in various fields of science. This study used FA to obtain the SD of questionnaire survey and to obtain the latent factor and factor loading of CARs.

The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

208

(1) Selection of representative product sample
Mobile phones are some of the most popular items in daily use; thus, this study chose the mobile phone as its experimental product. In the initial stage of research, 120 pictures of the most popular mobile phone brands

uced using 22 adjectival pairs and 60 mobile phone pictures. Forty participants: eight salespeople who are familiar with mobile phones, twelve consumers who are general users, fifteen college students from the department of design, and five mobile phone product designers were invited

were collected. In order to avoid any interference to a subject’s perception, the backgrounds of the 120 pictures were removed, and a standard size and

to partake in a survey experiment concerning the psychological images projected by mobile phone products. A user-friendly questionnaire interface

number were assigned and administered. Six mobile phone salespeople and four mobile phone designers were invited to select 60 representative pictures for the final statistical analysis.

was designed to collect the evaluation data in a more effective way. The order of presentation of the products were randomized to avoid any systematic effects, and on a scale between -1 to
+1.

(2) Selection of initial adjective pair
(4) Result of Factor analysis
In product design research, either single or pair wise adjectives can be used to describe consumers’ affective dimensions; pair wise adjectives are more suitable for describing consumers’ affective dimensions (Han & Hong,
2003). Common adjectives can be obtained from numerous sources in the literature. For analysis, we chose 22 pairs of adjectives that describe mobile phones from Chuang (Chuang, Chang
& Hsu, 2001), as shown in Table 1.
These were used as adjectival pairs in the following experiment.
(3)

Based on FA, using image survey data of the previous experiment and according to the principle component analysis results, a maximum variance of orthogonal rotation was adopted.
When the eigenvalue is greater than 1, three factors are obtained with a total loading value of explained variance of
86. 89%. The KMO value is 0. 897(>0.
8), which shows it is appropriate for FA.
Meantime, for an examination among the outliners of every adjective, data show the values are greater than 0. 5;

questionnaire

and, there is no double loading phenomenon in the component matrix after rotation, indicating that there are

First, a questionnaire was prod-

no outliners, as shown in Table 2.
Cronbach's Alpha value is 0. 92(>0. 5),

Experiment design and

The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

209

indicating considerable consistency in

coefficient is very low from the first to

the reliability of every adjective in every factor. Meanwhile, the SMC value is greater than 0. 5, which shows an extremely good contracted validity

the eleventh stage, which can be ignored. While we select the last ten stages (1~10 groups) in the twelfth to the twenty-first stages for analysis, the

for every adjective. We extracted three factors from FA, the three factors were
Evaluation, Potency, Activity (Osgood,

coefficient increase for the next stage will be calculated using the combination coefficient of every stage as

1957).

shown in Table 4.

The analysis process and result for cluster analysis

In Table 4, when four groups were combined into three groups at the

Representative CARs still could not be obtained based on the results of

fourth stage we found that the percentage increase of the combination coefficient was 77.94%, indicating an

FA. Therefore, this study further performed a cluster analysis on the factor loadings of FA. CA was used to assemble variables with similarity in the

increasing trend. Therefore, we can determine the three groups is the best number of clusters.

same cluster according to logic procedure and the similarity and dissimilarity of variables. The homogeneity of the variables in the same

(2) K-means clustering method

cluster was high, while the heterogeneity of variables in different clusters was high.

best cluster number was three clusters.
The K-means clustering method was be implemented two-stage cluster analysis, the result as shown in Table 5. In Table
5, the bold underlined numbers indicate the shortest distance from adjectives to

(1) Hierarchical cluster analysis
After FA, we use the HCA shown in the result of the Walter Method,

From the calculate result of combination coefficient increase ratio, the

the seed point, the three adjective (1)

indicating the cluster condition of every adjective without obtaining enough clusters. To obtain the required clusters, we analyze every adjective in the cluster analysis process. There are 21

The Procrustes analysis process
After obtaining the factor loadings from FA, the backward elimination process using PA is conducted to analyze

stages in the combination process, as shown in Table 3. The combination

the importance of the adjectives; their ranking can also be determined. The

The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

210

relative importance of each adjective

Table 1. The initial affective dimensions. adj. No.

The initial affective dimensions

adj. No.

The initial affective dimensions

1

Traditional-Modern

12

Unoriginal-Creative

2
3

Hard-Soft
Old-New

13
14

Simple-Complicated
Conservative-Avant-garde

4

Heavy-Handy

15

Standard-Outstanding

5
6

Obedient-Rebellious
Nostalgic-Futuristic

16
17

Common-Particular
Plain-Luxurious

7

Coarse-Delicate

18

Decorative-Practical

8
9

Masculine-Feminine
Rational-Emotional

19
20

Inert-Active
Personal-Professional

10
11

Hand Made-Hi Tech
Childish-Mature

21
22

Obtuse-Brilliant
Discordant-Harmonious

Table 2. The result of Factor analysis.
Factor

Affective dimensions communality

Cronbach’s
Alpha

SMC

1

2

3

Old

0.969

0.057

0.085

0.949

0.908

0.964

Standard

0.964

0.106

0.137

0.958

0.908

0.980

Common

0.956

-0.180

-0.054

0.950

0.909

0.975

Traditional

0.956

0.119

0.046

0.930

0.908

0.961

Conservative

0.953

-0.115

-0.121

0.935

0.910

0.961

Unoriginal

0.938

-0.130

-0.103

0.908

0.910

0.950

Plain

0.923

0.063

-0.088

0.863

0.911

0.911

Nostalgic

0.920

0.193

0.068

0.888

0.909

0.952

Obtuse

0.848

0.275

0.400

0.955

0.910

Inert

0.782

-0.456

0.967 adj. dimensions 0.893 No. adj. 0.173 No. The initial affective0.914
0.850

The initial affective dimensions

Coarse

0.782

0.287

0.466 1

Unoriginal-Creative

Personal

0.303

0.887

0.073

Hard

0.075

-0.854

0.251

Rational

0.224

-0.842

0.378

Decorative

-0.301

Childish

0.426

Hand made

0.569

Discordant

0.322

Obedient

0.373

Heavy

0.254

Simple
Masculine

0.573
0.095

Traditional-Modern
0.911
0.910

0.95212

2
3

Hard-Soft
0.884
Old-New

4

Heavy-Handy

0.921

0.908

0.798

0.926

0.837

0.903

0.924

0.924

13
14
15

Simple-Complicated
Conservative-Avant-garde
Standard-Outstanding

0.815

5
0.276 6

Obedient-Rebellious
0.831
0.931
Nostalgic-Futuristic

16
0.87217

0.757

0.341 7

0.871
Coarse-Delicate

0.917

0.92518

Decorative-Practical

0.712

0.119 8

0.845
0.916
Masculine-Feminine

0.89719

Inert-Active

0.277

0.833 9
10
-0.802

Rational-Emotional
0.874
0.918
Hand Made-Hi Tech
0.783
0.923

0.95020

0.005

0.78721

Personal-Professional
Obtuse-Brilliant

-0.126

11
0.738

Childish-Mature 0.920
0.625

0.73822

Discordant-Harmonious

0.161

-0.651

-0.638

0.643

0.778
0.830

0.920
0.928

Common-Particular
Plain-Luxurious

0.834
0.885

Final statistics
Eigenvalue

10.735

4.904

3.476

Variance(%)

47.775

22.702

16.415

Cumulative(%)

47.775

70.477

86.892

KMO

0.897

The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

211

Table 3. Cluster combined and coefficients stag cluster combined

e

cluster1

1
2

5
1

6
2

3
4

4
9

5
6
7

coefficients

cluster2

stage cluster first appears

next stage

cluster1

cluster2

0.000
0.003

0
0

0
0

5
6

8
11

0.007
0.011

0
0

0
0

6
17

3
1
13

5
4
14

0.016
0.025
0.044

0
2
0

1
3
0

14
8
13

8
9

1
16

7
17

0.071
0.107

6
0

0
0

15
11

10
11

19
12

21
16

0.150
0.207

0
0

0
9

18
16

12
13
14

18
13
3

20
22
10

0.296
0.399
0.547

0
7
5

0
0
0

17
20
15

15
16

1
12

3
15

0.802
1.213

8
11

14
0

18
19

17
18

9
1

18
19

1.658
2.924

4
15

12
10

19
20

19
20

9
1

12
13

4.193
7.461

17
18

16
13

21
21

21

1

9

10.996

20

19

0

Table 4. The combined coefficients increase ratio.
Increase ratio of the
Coefficient
coefficient
10

0. 296

34. 80%

9

0. 399

37. 09%

8

0. 547

46. 62%

7

0. 802

51. 25%

6

1. 213

36. 69%

5

1. 658

76. 36%

4

2. 924

43. 40%

3

4. 193

77. 94%

2

7. 461

47. 38%

1

10. 996

can be examined according to the calculated RSSDs values during the elimination process. In this study, a total of 22 adjective pairs, the deletion

stages. For example, the RSSDs values calculated in Steps 1, 9, 15 and 21 are shown in Figure 2(a), 2(b), 2(c), 2(d).

process is divided into 21 steps, due to space limitations, here list only four

In Figure 2., x-axis shown the 22 adjective pairs, y-axis shown the RSSD

The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

212

value size, from each stage and each

important, however, in Step 15, shown

adjective pair of rectangular diagram, we can see the RSSD value changes in circumstances. in Figure 2(c), adj9 is the most important, but the importance of adj19 exceeds that of adj8. This elimination process of adjectives still remains very

For Step 1, shown in Figure 2(a), the RSSD value of adj. 8 (Masculine-Feminine) is the largest which

stable even when the number of variables is small. In Step21, shown in
Figure 2(d), the final step was adj. 14、

implies that if it were to be eliminated from the adjective subset, the loss of information would be the greatest compared to the other adjectives. The

adj. 8. We observe the final step of the removal process, it has continued to maintain the overall structural.

adj. 9 (Rational-Emotional) gave the second greatest RSSD value, which means that it is of the second most

(2) Results of the adjective ranking

importance. In Step 9, shown in Figure
2(b), adj. 8 and adj. 9 remains the most

The results of the final adjective ranking obtained by the backward elimination process of PA are shown in
Table 6.

.
Table 5. The result of K-means cluster analysis. cluster adj. No.

affective dimensions
Seed point

0.300

0.820

0.28

0

1

Traditional-modern

0.961

0.003

-0.076

0.177

2

Hard-soft

-0.01

0.881

0.148

0.224

3
4

Old-new
Heavy-handy

0.970
0.325

0.070
0.244

-0.046
0.678

0.240
0.166

5

Obedient-rebellious

0.267

-0.058

-0.842

0.202

6
7

Nostalgic-futuristic
Coarse-delicate

0.938
0.865

-0.072
-0.133

-0.041
0.380

0.199
0.103

1

*

8

9
10

X

Y

Z

The distance

Masculine-feminine

0.088

0.718

0.553

0.208

Rational-emotional
Hand made-hi-tech

0.155
0.670

0.903
-0.619

0.253
0.111

0.499
0.531

11

Childish-mature

0.564

-0.654

0.355

0.566

19
21

Inert-active
Obtuse-brilliant

0.729
0.921

0.564
-0.121

0.013
0.304

0.900
0.685

Seed point

0. 57

0.16

-0.65

0

12

15

Unoriginal-creative
Standard-outstanding

0.892
0.978

0.228
0.027

-0.246
0.011

0.169
0.556

16
17

Common-particular
Plain-luxurious

0.909
0.905

0.285
0.037

-0.206
-0.209

0.228
0.340

13

Seed point
Simple-complicated

0.10
0.503

-0.64
-0.169

0.64
-0.704

0
0.538

14

Conservative-avant-grad

0.906

0.212

-0.264

0.449

18

e
Decorative-practical

-0.152

-0.806

0.399

0.772

20
22

Personal-professional

0.426

-0.829

0.122

0.359

Discordant-harmonious

0.458

-0.133

0.804

0.236

2

*

3

*
*
*
*

The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

213

(a) Step 1

(b) Step 9

(b) Step 15

(d)Step 21

Figure 2. The RSSDs value of adjectives calculated in (a)Step 1, (b)Step 9, (c)Step 15, (d) Step 21

Table 6. Results of adjectives ranking using PA
Rank
1

Affective dimension

Rank

Affective dimension

12

17

Masculine-feminine

13

4

3

12 Unoriginal-creative

14

18

4

16 Standard-outstanding

15

10

Hand made-hi tech

Rational-emotional

16

15

Standard-outstanding

6

13 Simple-complicated

17

20

Personal-professional

7

19 Inert-active

2

5

14 Conservative-avant grade
8

9

Plain-luxurious
Heavy-handy
Decorative-practical

18

6

8

5

Obedient-rebellious

19

22

Discordant-harmonious

9

3

Old-new

20

21

Obtuse-brilliant

10

2

Hard-soft

21

11

Childish-mature

11

1

Traditional-modern

22

7

Coarse-delicate

During the reduction of adjectives at each stage, RSSD value could be calculated. Therefore, PA approach provided a standard method for assessing information loss. The results of the final calculation also determined

Nostalgic-futuristic

the priority order of variables of adjectives. The comparison for CA and PA
The three adjective pairs selected by CA were Coarse-Delicate, Unor-

The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

214

iginal -Creative, Discordant -Harmon-

For comparing the differences of

ious; the adjectives selected by PA were
Masculine - Feminine, Rational- Emotional, Unoriginal -Creative, Conservative - Avant - garde and Common -

CA and PA, the distribution of all adjectives were marked in the space of three extracted factors in Figure 3(a) and 3(b). In order to examine the

Particular. It is important to note that only Unoriginal-Creative are the same, and the remaining pairs of adjectives

relationship between the selected adjectives and the Osgood’s three factors, the areas of factors Evaluation, Potency,

chosen by the two methods are different. The results show that there were differences between these two methods.

and Activity are drawn in red, green, and blue, respectively.

(a)

(b)

Figure 3. The adjectives exhibited against (a) factor 1-factor 2 and (b) factor 3-factor 2

As shown in Figure 3(a), the top five adjectives selected using PA are marked in solid circle. The last seven-

structure of the initial affective dimensions. That is, the distribution of these five adjectives exhibit similar global

teen adjectives in the ranking are marked by an empty circle. Among the top five adjectives, there are three ad-

structures compared to all the other adjectives. jectives (adj. 12, adj. 14 and adj. 16) picked from the Evaluation factor, one adjectives (adj. 9) from the Potency factor, and one adjective (adj. 8) from

Figure 3(b) shows the distribution of all adjectives in the factor space for two-stage CA. The adjectives in clusters 1 to 3 are marked in circle, triangle

the Activity factor. This is the clearest evidence that the proposed PA approach is capable to preserve the global

and square, respectively. The top four selected adjectives, which are closest to each centroid of the clusters, are drawn

The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

215

in solid marks. In Figure 3(b), it can be

hasizes maintaining the overall struct-

seen that there are two adjectives (adj.
7 and adj. 12) picked from the
Evaluation factor, one adjective (adj. 22) from the Activity factor. However,

ure of the data, it cannot be analyzed based on the local relationship between the adjectives so the natures of the two methods are completely different.

there is not one adjective selected from the Potency factor. The result of CA is failure to select the representative

When using CA, the distance are very close to some adjective and seed point each other, this is not easy to determine

adjective from the Potency factor. This is due to that the two groups of the
Potency factor are located far from each other.

whether which one is the most appropriate adjective. Therefore, to extract adjective is still difficult using
CA, for example, in figure 3 (b), the adj.

Conclusions

12 and adj. 16 are very close to each other on the group's seed point.

The research results show that there are some differences in the best affective dimensions obtained by the
PA and CA methods. The CA method

PA can be used to analyze the proper process for deleting the adjectives for which their influences to the overall structure is minimized. It

puts emphasis on classifying adjectives with similar factor loading into the same clustering according to the relationship between them. While PA emp-

provides researchers with a “global” analyzing tool in conjunction with the
“local” analyzing tool of CA.

Reference

Andrade, J. M., Gomez-Carracedo, M. P.,
Krzanowski, W., & Kubista, M.

Alcantara, E., Artacho, M. A., Gonzalez,
J. C., & Garcia, A. C. (2005). Ap-

(2004). Procrustes rotation in anal-

plication of product semantics to footwear design. Part I—Identification of footwear semantic

metrics and Intelligent Labor- atory

space applying differential semantics. International Journal of
Industrial
Ergonomics,
35,
713-725.

Chen, D., Wang, J., Chen, X. & Xu, X.

ytical chemistry, a tutorial. ChemoSystems, 72, 123-132.

(2010). A Search algorithm for clusters in a network or graph. International Journal of Digital Content
Technology and its Applications, vol. 4, no. 6, 115-122.

The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

216

Chuang, M. C., Chang, C. C., & Hsu, S.

Osgood, C. E., Suci, C. J., & Tannen-

H. (2001). Perceptual factors under-

baum, P. H. (1957). The measure-

lying user preferences toward pro-

ment of meaning. Champaign, IL:

duct form of mobile phones. Inter-

University of Illinois Press.

national Journal of Industrial Ergonomics, 27, 247-258.

Sahmer, K. & Qannari, E. M. (2008).
Procedures for the selection of a

Han, S. H., & Hong, S. W. (2003). A

subset of attributes in sensory pro=

systematic approach for coupling

filing. Food Quality and Preference,

user satisfaction with product de-

vol. 19, 141-145.

sign. Ergonomics, 46(13/14), 14411461.

Salvador, M., Pedro, C. & Margarita, V.
(2005). Semantic differential app-

Hurley, J. R. and R. B, Cattell. (1962).

lied to the evaluation of machine

The Procrustes Program: Producing

tool des- ign. International Journal

Direct Rotation to Test a Hypothe-

of Industrial Ergonomics, vol. 35,

sized Factor Structure. Computers

1021-1029.

In Behavioral Science, 7: 258-262.
Vigneau, E., & Qannari, E. M. (2003).
Hsu, S. H., Chuang, M. C., & Chang, C.

Clustering of variables around latent

C. (2000). A semantic differential

components. Communications in

study of designers' and users' prod-

Statistics Simulation and Comp-

uct form perception. International

utation 32(4), 1131-1150.

Journal of Industrial Ergonomics,
25, 375-391.
Jindo, T., Hirasago, K., & Nagamachi, M.
(1995). Development of a design sup- port system for office chairs using 3-D graphics. Inter- national
Journal of Industrial Ergonomics,
15, 49-62.
Krzanowski, W. J. (1987). Selection of variables to preserve multivariate data structure, using principal components. Applied Statistics, 36(1),
22-33.
The International Journal of Organizational Innovation Vol 8 Num 1 July 2015

217

Copyright of International Journal of Organizational Innovation is the property of Frederick
L. Dembowski and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.

Similar Documents