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. 