latent semantic engineering – a new conceptual user-centered design approach

Advanced Engineering Informatics 26 (2012) 456–473

Contents lists available at SciVerse ScienceDirect

Advanced Engineering Informatics

journal homepage: www.elsevier .com/ locate /ae i

Latent Semantic Engineering – A new conceptual user-centered design approach

Gregory C. Smith, Shana Smith ⇑Department of Mechanical Engineering, National Taiwan University, Taiwan, ROC

a r t i c l e i n f o

Article history:Received 14 August 2011Received in revised form 20 January 2012Accepted 24 February 2012Available online 21 March 2012

Keywords:Conceptual designUser-centered designLatent Semantic Engineering

1474-0346/$ - see front matter � 2012 Elsevier Ltd. Adoi:10.1016/j.aei.2012.02.012

⇑ Corresponding author. Address: No.1, Sec. 4, RooROC. Tel.: +886 2 33662692; fax: +886 2 23631755.

E-mail address: [email protected] (S. Smith).

a b s t r a c t

User-centered design (UCD) plays a vital role in the product development process. UCD approachesmatch designs to user needs. Matching designs to needs improves product quality, customer satisfaction,and product success. The goal of this study is to improve matching accuracy. To achieve the goal, thisstudy introduces conceptual UCD and Latent Semantic Engineering (LSE), a new conceptual UCDapproach, defines measures for model accuracy (MA), conceptual matching accuracy (CMA), user-cen-tered design accuracy (UCDA), and conceptual design accuracy (CDA), and compares the LSE approachto other approaches. The LSE approach models conceptual design processes more accurately than otherapproaches. Functional approaches use subjective weights to match functional needs to physical designs.Emotional (aesthetic) approaches use statistical models to match physical designs to emotional needs.The LSE approach uses a LSE semantic space model to create complete conceptual (aspirational, emo-tional, functional, physical) designs from complete conceptual needs. The LSE approach improves modelaccuracy (MA) and matching accuracy (CMA, UCDA, and CDA), compared to other approaches. The LSEapproach creates design descriptions more accurately than other approaches. The LSE approach matchesdesigns to user choices, needs, and concepts more accurately than other approaches. This study uses theLSE approach to create customized cell phone designs for individual users. The approach can also be usedto create other conceptual designs for either individual or multiple users. Study results can be used toimprove product quality, customer satisfaction, product success, and the product development process.

� 2012 Elsevier Ltd. All rights reserved.

1. Introduction

User-centered design (UCD) plays a vital role in the productdevelopment process. In today’s market, product success is deter-mined by customer satisfaction [1]. Customer satisfaction isachieved by meeting customer (user) needs. UCD approachesmatch designs to user needs [2]. Matching designs to needs im-proves product quality, customer satisfaction, and product success.

However, matching designs to needs is inherently difficult andinaccurate [3,4]. Customers generally use non-technical words todescribe their needs or wants [5]. UCD approaches must matchtechnical design descriptions to non-technical words (needs).Matching accuracy is generally very low [6,7]. As a result, mostproducts still do not match customer (user) needs [8].

The goal of this study is to improve matching accuracy. Toachieve the goal, this study introduces conceptual UCD and LatentSemantic Engineering (LSE), a new conceptual UCD approach, de-fines measures for model accuracy and matching accuracy, andcompares the LSE approach to other approaches. This study uses

ll rights reserved.

sevelt Road, Taipei, Taiwan,

a case study to compare the LSE approach to two KE (Kansei engi-neering) approaches.

1.1. Conceptual user-centered design

In today’s market, products must meet customer (user) needs atfour conceptual levels. At the aspirational level, products must en-hance personal image or status. At the emotional level, productsmust elicit positive feelings or emotions. At the functional level,products must provide adequate utility or functionality [9]. Atthe physical level, products must include specific aesthetic or func-tional design elements [10].

The LSE approach models conceptual design (CD) processesmore accurately than other approaches. Functional approachesmatch functional needs to functional or physical designs [3]. Emo-tional (aesthetic) approaches match physical (aesthetic) designs toemotional needs [2]. The LSE approach creates complete concep-tual (aspirational, emotional, functional, physical) designs fromconceptual needs.

Complete designs meet user needs at all four conceptual levels.As a result, they are more successful than other products. Forexample, consumers feel that Apple’s computer products are bothemotionally ‘appealing’ and functionally ‘easy to use’. As a result,Apple’s iPod is the best selling MP3 player, worldwide. Apple’s

http://dx.doi.org/10.1016/j.aei.2012.02.012

mailto:[email protected]

http://dx.doi.org/10.1016/j.aei.2012.02.012

http://www.sciencedirect.com/science/journal/14740346

http://www.elsevier.com/locate/aei

G.C. Smith, S. Smith / Advanced Engineering Informatics 26 (2012) 456–473 457

iMac is also more popular than many less expensive products[1,11].

1.2. Latent Semantic Engineering

In today’s market, customers use non-technical words to de-scribe what they need or want [5]. At the (aspirational, emotional,functional, physical) levels, customers want (‘prestigious’, ‘excit-ing’, ‘easy to use’, ‘blue’ ‘camera’) phones. UCD approaches mustmatch technical design descriptions (‘type’, ‘screen shape’, ‘bodyshape’, ‘number keys’, ‘function keys’) to non-technical words(needs).

Functional approaches use subjective weights or statisticalmodels to match needs to designs [3]. Emotional approaches usestatistical models to match designs to needs [2,4]. The LSE concep-tual approach uses LSE semantic space models to create designsfrom needs. The LSE conceptual semantic space approach improvesmodel accuracy and matching accuracy, compared to otherapproaches.

The LSE approach creates LSE conceptual semantic space mod-els that are different than other semantic space models. LSA (latentsemantic analysis) approaches create LSA models that contain onelevel [12–14]. The KE-LSA (Kansei engineering-LSA) approach cre-ates KE-LSA models that contain one level. The LSE approach cre-ates LSE models that contain many conceptual levels.

The LSE approach uses LSE models to model different processesthan other approaches. LSA approaches use LSA models to modelmemory, learning, and document indexing processes [12–14].The KE-LSA approach uses KE-LSA models to model KE processes.The LSE approach uses LSE models to model CD processes. All ofthe approaches use semantic space models to improve modeland matching accuracy.

The LSE approach uses LSE semantic space models in differentways than other approaches. LSA approaches use LSA models tomatch keywords to documents [12–14]. The KE-LSA approach usesKE-QT1 (KE-Quantification Theory Type 1) and KE-LSA models tomatch physical designs to emotional needs. The LSE approach usesLSE models to create new conceptual designs from conceptualneeds.

1.3. Measures

The LSE approach improves model accuracy (MA) and matchingaccuracy (CMA, UCDA, and CDA), compared to other approaches.Model accuracy (MA) measures accuracy for creating designdescriptions. Conceptual matching accuracy (CMA), user-centereddesign accuracy (UCDA), and conceptual design accuracy (CDA)measure accuracy for matching designs to user choices, needs,and concepts.

1.4. Case study

This study uses a case study to compare the LSE approach to KE-QT1 and KE-LSA approaches. The LSE approach can create concep-tual designs that contain any combination of conceptual designelements, from any combination of conceptual needs. KE-QT1and KE-LSA approaches match physical designs to emotionalneeds. Therefore, the case study uses the LSE approach to createconceptual (emotional, physical) designs from conceptual (emo-tional, physical) needs.

The case study compares the LSE approach to KE approaches. KEapproaches create aesthetic (visual) designs [2]. Therefore, the casestudy uses the LSE approach to create aesthetic (visual) designs. Tocreate individually customized designs, a UCD approach must cre-ate a wide variety of designs, with high accuracy. Therefore, the

case study uses the LSE approach to create individually customizeddesigns.

Cell phones are single-user products, cell phones are the mostpopular product in the world, cell phone designs vary widely,and cell phone markets are customer driven and competitive[15]. Traditional cell phones have more visual design elementsthan smart phones, handheld computers, or PDAs. Therefore, thecase study uses the LSE approach to create traditional cell phonedesigns.

The LSE approach can also be used to create other conceptualdesigns for either individual or multiple users. For example, theLSE approach can be used to create screen functions for smartphones, handheld computers, or PDAs, from any combination ofconceptual needs. For multiple users, the LSE approach can averageneeds or create individual designs and choose the most commondesigns.

1.5. Outline

Section 2 describes prior UCD approaches. Section 3 describesprior studies that tried to improve matching accuracy. Section 4describes the research goal of this study. Section 5 describes thecase study. Section 6 describes the KE-QT1 approach. Section 7 de-scribes the LSA approach. Section 8 describes the KE-LSA approach.Section 9 describes the LSE approach. Section 10 compares casestudy design results.

2. Prior approaches

In early markets, product success was determined by producergoals. Producer goals were achieved by meeting design goals.Designers used producer-centered design (PCD) approaches tomatch designs to design goals. Matching designs to design goalsimproved product success, with respect to producer goals. Match-ing designs to design goals did not improve product success, withrespect to user needs.

For example, producers set goals to ‘reduce cost’. Designers setgoals to reduce individual component costs. Designers used PCDapproaches to match ‘low cost’ designs to design goals. Matchingdesigns to design goals improved product success with respect toproducer ‘cost’ goals. Matching designs to design goals did not im-prove product success with respect to, for example, user ‘perfor-mance’ needs.

In today’s market, product success is determined by customersatisfaction. Customer satisfaction is achieved by meeting cus-tomer (user) needs. Designers use UCD approaches to match de-signs to user needs [15]. For example, users provide weightedneeds for ‘performance’ (7) and ‘price’ (4). Designers use UCD ap-proaches to create ‘high performance’ designs that match user ‘per-formance’ needs.

Designers use three UCD approaches to match designs to userneeds. Designers use the quality function deployment (QFD) ap-proach to match functional needs to physical designs [16,17], theconjoint analysis (CA) approach to match physical designs to userrankings [4], and the KE approach to match physical designs toemotional needs [2]. The QFD, CA, and KE approaches were intro-duced in the 1960s, 1970s, and 1980s.

2.1. Quality function deployment

In the QFD approach, customers provide functional require-ments, with real-value weights, for example (1–7). Designers usesubjective ordinal-value weights, for example (1, 3, 9), to matchcustomer functional requirements to physical design elements.Designers use simple calculations to combine weights and choose

Table 1Prior studies.

Study Approach Product Results

Hsu et al. [16] QFD Cell phone Needs – ‘call quality’ (3.9), ‘modeling’ (3.9)QFD-Kano Cell phone Needs – ‘call quality’ (4.6), ‘modeling’ (4.1)

Kannan [3] QFD Auto part Design – ‘conformance’ (1.0), ‘features’ (0.9)QFD-fuzzy Auto part Design – ‘conformance’ (0.7, 1, 1), ‘features’ (0.6, 0.9, 0.9)

Ling et al. [17] CA-element Cell phone Design – ‘appearance’ (5.6), ‘size’ (5.6), ‘menu’ (5.6)CA-design Cell phone Design – ‘appearance’ (0.26), ‘size’ (0.09), ‘menu’ (0.06)

Barone et al. [4] CA-KE Cell phone Design – ‘dimensions’ (0.51), ‘camera’ (0.11), ‘antenna’ (0.10)

Chen and Chuang [1] KE-grey cell phone Design – ‘top shape’ (0.79), ‘body shape’ (0.75), ‘outline’ (0.70)

Chen et al. [18] KE-culture Cell phone Design – ‘body’, ‘screen’, ‘panel’, ‘buttons’

Seva and Helander [10] KE-aesthetic Cell phone Design – ‘number of colors’ (–0.23,–0.29,–0.20)KE-function Cell phone Design – ‘screen area’ (–0.39, –0.55, –0.56)

Lai et al. [6] KE-NN Cell phone Reverse – design – ‘color’ (0.62), ‘size’ (0.23), accuracy (50%)KE-NN Cell phone forward – design – ‘color’ and ‘size’

Fukushima et al. [7] KE Copier color Reverse – design – ‘lightness’ (0.68), ‘chroma’ (0.54), ‘hue’ (0.53)KE-math Copier color Forward – design – improve ‘desirability’ (6%)

Nagamachi [19] KE Auto part Reverse – design – calculate Kansei valuesKE Auto part Forward – design – ‘shape’

Smith and Smith KE-QT1 Cell phone Design – CMA (7.6%), UCDA (83.5%), CDA (64.1%)KE-LSA Cell phone Design – CMA (7.6%), UCDA (83.5%), CDA (64.1%)LSE-M Cell phone Design – CMA (15.2%), UCDA (86.9%), CDA (68.4%)LSE-D Cell phone Design – CMA (36.7%), UCDA (89.6%), CDA (73.0%)

458 G.C. Smith, S. Smith / Advanced Engineering Informatics 26 (2012) 456–473

design elements. Designers use the design elements to create anew design.

Designers have used the QFD approach, with some degree ofsuccess, since the 1960s. However, the QFD approach is inherentlyinaccurate [3,17]. Customers use subjective, inaccurate weights torank requirements. Designers use subjective, inaccurate weights tomatch requirements to design elements [17]. Designers combinethe weights to choose design elements. The new design may notmatch requirements.

2.2. Conjoint analysis

In the CA approach, users do not provide functional require-ments. Designers choose physical design elements, create sampledesigns (virtual prototypes, physical prototypes, or actual prod-ucts), and show the sample designs to users. Users rank the sam-ple designs. Designers use regression analysis to analyze therankings and choose physical design elements. Designers createa new design [4].

The CA approach uses ranking and statistical matching to im-prove matching accuracy. However, the CA approach is still inher-ently inaccurate. The sample designs may not match requirements.The sample design rankings are not detailed. Designers use thesample designs and the sample design rankings to choose physicaldesign elements. As a result, the new design may not matchrequirements.

2.3. Kansei engineering

In the KE approach, users do not provide emotional needs.Designers choose physical design elements, create sample designs,and show the sample designs to users. Users rank the sample de-signs with respect to several emotional (Kansei) words. Designersuse Quantification Theory Type 1 (QT1) regression analysis to ana-lyze the rankings and choose physical design elements. Designerscreate a new design [2].

The KE approach uses detailed Kansei word rankings to improveranking accuracy. However, the KE approach is still inherentlyinaccurate. The sample designs may not match user needs. The

sample design rankings are not user needs. Designers use the sam-ple designs and the sample design rankings to choose physical de-sign elements. As a result, the new design may not match userneeds.

2.4. Reverse Kansei engineering

In the reverse Kansei engineering (KE-QT1) approach, usersprovide emotional (Kansei value) user needs. Designers use KE-QT1 values from sample designs to create a KE-QT1 model, usethe KE-QT1 model to calculate Kansei values for new designs,and use correlation to match Kansei values for new designs toKansei values for user needs. In the iterative KE-QT1 approach,the process repeats [2].

The KE-QT1 approach uses actual user needs to improve match-ing accuracy. The iterative KE-QT1 approach uses iterative reverseengineering to improve matching accuracy. However, both ap-proaches are still inherently inaccurate. The KE-QT1 model maynot calculate Kansei values accurately. Correlation may not matchKansei values accurately. As a result, the new design may notmatch user needs.

2.5. Forward Kansei engineering

In the forward KE approach, users provide emotional (Kanseivalue) user needs. Designers use a KE-QT1 model with rules, expertsystems, neural networks, genetic algorithms, fuzzy logic, or math-ematical models, to create a forward KE model. Designers use theforward KE model to match Kansei values for user needs to physi-cal design elements. Designers use the design elements to create anew design [2].

The forward KE approach is more efficient than the iterativeKE-QT1 approach. The forward KE approach is less accurate thanthe iterative KE-QT1 approach. Overall accuracy depends uponreverse KE-QT1 model accuracy and forward KE model accuracy.Designers use the KE-QT1 model and the forward KE model tochoose design elements. As a result, the new design may notmatch user needs.


2.6. Survey techniques

Most user-centered design approaches use Likert-scale surveysto identify user needs. Most surveys use semantic differential (SD)scales to improve requirement ranking accuracy [11]. For example,a traditional survey might use a Likert-scale for ‘exciting’ thatranges from 1 (‘not’) to 7 (‘very’). An improved survey might usean SD scale that ranges from 1 (‘very exciting’) to 7 (‘very plain’).

Some approaches also use Kano models to improve ranking accu-racy [17]. Kano models classify user needs into five categories: (1)‘attractive’ needs increase satisfaction, (2) ‘reverse’ needs decreasesatisfaction, (3) ‘must-be’ needs decrease satisfaction, when theyare not met, (4) ‘one-dimensional’ needs have a linear relationshipwith satisfaction, and (5) ‘indifferent’ needs do not affectsatisfaction.

3. Prior studies

Table 1 shows prior studies that tried to improve UCD matchingaccuracy. Most combined two or more traditional UCD approaches.Some combined traditional UCD approaches with fuzzy logic, greyrelational analysis, neural network (NN), or mathematical model-ing techniques. None of the approaches improved matching accu-racy, significantly, compared to existing approaches.

Hsu et al. [16] used a QFD approach and a QFD-Kano model ap-proach to rank functional requirements, for the top three cellphone brands in Taiwan. The QFD approach ranked ‘call quality’(3.9) and ‘modeling’ (3.9) equally. The QFD-Kano model approachranked ‘call quality’ (4.6) higher than ‘modeling’ (4.1). The ap-proach used a Kano model to improve QFD requirement rankingaccuracy (requirement weights). The approach did not improveQFD matching accuracy (matching weights).

Kannan [3] used a QFD approach and a QFD-fuzzy logic ap-proach to match functional requirements to functional design ele-ments, for an automobile fork assembly. The QFD approach rankedfork assembly ‘conformance’ (1.0) higher than ‘features’ (0.9). TheQFD-fuzzy logic approach showed that ‘features’ (0.6, 0.9, 0.9)might be more important than ‘conformance’ (0.7, 1, 1). The ap-proach increased QFD matching uncertainty. The approach couldreduce or increase QFD matching accuracy.

Ling et al. [17] used two CA approaches to match cell phone‘appearance’, ‘size’, and ‘menu’ to customer satisfaction. In a CA-element approach, users ranked design elements (5.6), (5.6),(5.6). In a CA-design approach, users ranked designs. Designersused regression analysis to rank design elements (0.26), (0.09),(0.06). The study showed that statistical matching techniques aremore accurate than subjective matching techniques. The approachdid not improve CA matching accuracy.

Barone et al. [4] used a combined CA-KE approach to match sixcell phone design elements to four Kansei words. Users ranked sixdesign elements and eight sample designs. Designers usedweighted regression analysis, with weights derived from the de-sign element rankings, to match ‘dimensions’ (0.51), ‘camera’(0.11), and ‘antenna’ (0.10) to the Kansei word ‘appealing’. The ap-proach used sample design rankings and statistical matching tech-niques. The study did not determine matching accuracy.

Chen and Chuang [1] used a KE-Kano approach to match ten de-sign elements to six Kansei words. The approach used grey relationalanalysis to determine that ‘top shape’ (0.79), ‘body shape’ (0.75), and‘outline division style’ (0.70) had the greatest impacts on multipleKansei words. The approach matched design elements to multipleKansei words. The approach did not improve KE-Kano matchingaccuracy.

Chen et al. [18] used a KE approach, with four bilingual five-point Likert-scale surveys, to determine that many Kansei wordsfor cell phones are similar in different countries, while some are

unique to a single country. The study also determined that ‘body’,‘screen’, ‘panel’, and ‘buttons’ had the greatest impact on Kanseiwords, in all four countries. The approach used bilingual surveysto improve KE user need ranking accuracy. The approach did notimprove KE matching accuracy.

Seva and Helander [10] used a KE approach, with physical mod-els, to match ‘screen area’ to ‘amazed’ (�0.39), ‘contented’ (�0.55),and ‘hopeful’ (�0.56), in one country, and ‘number of colors’ to‘amazed’ (�0.23), ‘contented’ (�0.29), and ‘hopeful’ (�0.20), in an-other country. The approach used a KE approach to match func-tional and aesthetic physical design elements (‘screen area’ and‘number of colors’) to emotional user needs. The approach didnot improve KE model accuracy or KE matching accuracy.

Lai et al. [6] used KE, reverse KE-NN, and forward KE-NN ap-proaches to match cell phone color and size to user Kansei. TheKE approach determined that ‘color’ (0.62) had a stronger impacton user Kansei than ‘size’ (0.23). The reverse KE-NN approachmatched user Kansei word rankings for twelve new cell phone de-signs, with 50% accuracy. The forward KE-NN approach matchednew user needs to designs. The study did not determine forwardKE-NN accuracy or improve KE matching accuracy.

Fukushima et al. [7] used a reverse KE approach to match copier‘skin color’ to user ‘desirability’ rankings. They used a forward KEapproach with a mathematical model to improve ‘skin color’ ‘desir-ability’ by 6% (from 36% to 42%). The approach did not improvematching accuracy, over the reverse KE (KE-QT1) approach.

Nissan used a hybrid (reverse and forward) KE approach to de-sign a new steering wheel for a passenger car [19]. Designers useda forward KE approach to match new user needs to a candidatesteering wheel design, modified the design, and used a reverseKE approach to calculate Kansei values for the new design. The ap-proach did not improve matching accuracy, over the reverse KE(KE-QT1) approach.

4. Research goal

The goal of this study is to improve matching accuracy. Toachieve the goal, this study introduces conceptual UCD and theLSE approach, defines MA, CMA, UCDA, and CDA, and comparesthe LSE approach to KE-LSA and KE-QT1 approaches. This studyshows that the LSE approach improves model accuracy (MA) andmatching accuracy (CMA, UCDA, and CDA) compared to the KE-QT1 approach.

Prior studies show that the KE-QT1 approach improves modelaccuracy and matching accuracy, compared to other UCD ap-proaches. QFD approaches use subjective ranking and subjectivematching. CA and KE approaches use subjective ranking and statis-tical matching. KE-QT1 approaches use reverse engineering andstatistical matching. Forward KE approaches use forward and re-verse models.

Therefore, this study shows that the LSE approach improves mod-el accuracy (MA) and matching accuracy (CMA, UCDA, and CDA)compared to both the KE-QT1 approach and other UCD approaches.The LSE approach creates designs more accurately than other ap-proaches. The LSE approach matches designs to user choices, needs,and concepts more accurately than other approaches.

The LSE approach is also more versatile than other UCD ap-proaches. The LSE approach can match conceptual designs to con-ceptual needs or create new conceptual designs from conceptualneeds. Prior studies use QFD or CA approaches to match functionalor physical designs to functional needs. Prior studies use KE ap-proaches to match physical designs to emotional needs.

This study uses a LSE-D (design) approach to create conceptual(emotional, physical) designs from conceptual (emotional, physi-cal) needs. Prior studies use KE approaches to match physical de-signs to emotional needs. One study used a KE approach to

Table 2Case study.

Identify user needs

– Identify Kansei words – Create a survey – Complete the survey

Create sample design rankings

– Create sample designs – Create a survey – Complete the survey

Create user need descriptions

– Create Kansei values

Create a model

– Create design by KW matrix – Create a KE-QT1 model

Create design descriptions

– Choose designs – Calculate Kansei values – Create Kansei values – Match Kansei values – Choose design descriptions


– Create Kansei vectors

Create a model

– Create design by KW matrix – Create a KE-QT1 model – Calculate SVD of the model


– Choose designs– Calculate Kansei values – Create Kansei vectors – Project Kansei vectors – Match Kansei vectors – Choose design descriptions


– Create conceptual vectors

Create a model

– Create CDE by CD matrix – Calculate SVD of the matrix


– Project conceptual vectors – Match conceptual vectors – Choose CD descriptions

KE-QT1 approach KE-LSA approach LSE-D approach


match functional and aesthetic physical design elements (‘screenarea’ and ‘number of colors’) to emotional needs (‘amazed’, ‘con-tented’, and ‘hopeful’) [10].

This study uses a KE-QT1 approach and a KE-LSA approach tomatch physical designs to emotional needs, uses a LSE-D approachto create conceptual (emotional, physical) designs from conceptual(emotional, physical) needs, and determines, KE-QT1, KE-LSA, andLSE-D MA (87.4%, 87.4%, 100.0%), CMA (7.6%, 12.7%, 36.7%), UCDA(83.5%, 90.4%, 89.6%), and CDA (64.1%, 63.9%, 72.7%).

This study shows that the LSE-D approach improves MA andCMA, UCDA, and CDA, compared to the KE-QT1 approach. Priorstudies do not determine model accuracy (MA) for creating newdesign descriptions. One study determined KE-NN accuracy, forpredicting Kansei value rankings for new designs (50%) [6]. Priorstudies do not determine matching accuracy (CMA, UCDA, or CDA).

This study uses a case study to determine KE-QT1, KE-LSA, andLSE-D MA and CMA, UCDA, and CDA. The case study compares theLSE approach to KE approaches. Therefore, the case study uses theLSE approach to create conceptual (emotional, physical) designsfrom conceptual (emotional, physical) user needs. The approach

can also create complete conceptual designs from complete con-ceptual user needs.

5. The case study

The case study compares the LSE approach to KE-QT1 and KE-LSA approaches. The case study uses the KE-QT1, KE-LSA, and LSEapproaches to create customized cell phone designs for individualusers. In the case study, all three approaches use the same designprocess. The design process consists of three steps: create userneed descriptions, create a model, and create design descriptions(Table 2).

All three approaches use the same user needs to create userneed descriptions. All three approaches use the same sample de-sign rankings to create models. All three approaches use differenttechniques to create user need descriptions, models, and designdescriptions. The case study uses the user need descriptions, mod-els, and design descriptions to determine MA and CMA, UCDA, andCDA for all three approaches.

Table 3User needs.

1 4 7

Elegant h j h h h h h OrdinarySimple h h j h h h h ComplexHigh Tech h h h j h h h TraditionalLuxurious h h j h h h h BasicBeautiful h h j h h h h PlainUnique h h j h h h h Common


The case study uses matching (difference between values, inpercent) to determine MA and CDA. For each user, MA is between0% and 100%. CMA is either 0% or 100%. The case study uses vectorcosine matching (cosine of the angle between vectors, in percent)to determine UCDA and CDA. For each user, UCDA is between 0%and 100%. CDA is also between 0% and 100%.

5.1. Identify user needs

To identify Kansei words, designers used factor analysis to groupKansei words from magazines, advertisements, designers, andjournals into six representative Kansei word pairs: (1) ‘elegant–or-dinary’, (2) ‘simple–complex’, (3) ‘high tech–traditional’, (4) ‘luxuri-ous–basic’, (5) ‘beautiful–plain’, and (6) ‘unique–common’.

To identify user needs, designers used the six Kansei word pairsto create a two-part Internet survey. The first part contained six 7-point SD Likert scales, one SD scale for each of the Kansei wordpairs (Table 3). The survey asked users to provide Kansei value userneeds for a cell phone. Table 3 shows User 39’s response.

The second part contained 12 cell phone images (Table 4). Thesurvey asked each user to choose the closest match to their idealcell phone. Table 4 contains four vertical, four folding, and foursliding phones. Phones within each type represent various aes-thetic (visual) design elements (screen shapes, body shapes, andkey styles).

Seventy-nine users, with ages between 21 and 28, completedthe survey. Users had a significant impact on user needs, Kanseiwords had a significant impact on user needs, and phones had asignificant impact on user choices, (p-values = 0.0000). Table 4 re-sults show that most of the users (63.4%) chose Phone 3, Phone 5,or Phone 10. Overall, 24.1% of the users chose Phone 3. User 39chose Phone 5 (bold).

Table 4User choices.

The survey used traditional KE techniques to capture userneeds. Traditional KE techniques control user response (withKansei words and SD surveys) and user interaction with theproduct (with visual images), to capture emotional needs andemotional response to aesthetic (visual) design elements. The sur-vey also controlled age, to remove age bias from the results.

The survey did not control other factors. The survey did not usea Kano model to modify ranking results. The survey was non-selec-tive, within the target age range. Therefore, case study results canbe generalized to a random population of cell phone users, in thetarget age range, that respond to Internet surveys.

5.2. Create sample design rankings

To create sample designs, designers analyzed cell phones, ana-lyzed cell phone images, and identified aesthetic (visual) designelements. Table 5 results show that cell phones can be groupedinto three phone types: ‘vertical’, ‘folding’, and ‘sliding’. Resultsalso show that ‘phone type’, ‘screen shape’, ‘top shape’, ‘bodyshape’, ‘bottom shape’, ‘number keys’, and ‘function keys’ deter-mine overall visual appearance.

In Table 5, design elements and levels represent 37 = 2187 sampledesigns. Model size and sample design ranking time and effort de-pend upon the number of sample designs. To reduce model sizeand ranking time and effort, designers used a Taguchi L18 orthogonalarray to create a minimum spanning set of eighteen sample designs(Table 6). Designers used SolidWorks to create virtual prototypes.

To create sample design rankings, designers used the eighteensample designs, in Table 6, and the six Kansei word pairs, in Table 3,to create a new Internet survey. The survey contained one virtualprototype (Table 6) and one set of six Kansei word ranking scales(Table 3) for each of the 18 sample designs. The survey asked eachuser to rank each of the eighteen sample designs, with respect toeach of the six Kansei words.

The survey was open on the Internet until 52 users, 26 studentsand 26 working professionals, 26 male cell phone users and 26 fe-male cell phone users, with ages between 21 and 28, completed thesurvey. Sample designs had a significant impact on rankings(p-value = 0.0002). Kansei words had a significant impact on rank-ings (p-value = 0.0048). Table 7 results show that users rankedSamples 18, 1, 7, 7, 18, and 9 most ‘elegant’, ‘simple’, ‘high tech’,‘luxurious’, ‘beautiful’, and ‘unique’.

Table 5Design elements.

‘arc’

Element Level 1 Level 2 Level 3

‘phone type’

‘screen shape’

‘top shape’

‘body shape’

‘bottom shape’

‘number keys’

‘function keys’

‘vertical’ ‘folding’ ‘sliding’

‘round’ ‘arc’

‘1:1’ ‘4:3’

‘straight’

‘round’

‘round’ ‘arc’

‘straight’ ‘round’ ‘arc’

‘five keys’ ‘eight keys’‘seven keys’

’25:16’

‘straight’

‘straight’

Table 6Sample designs.

Sample 1 Sample 2 Sample 3







The study controlled age range, ratio of students to workingprofessionals, and ratio of male subjects to female subjects to re-move demographic bias from the rankings. The survey was non-selective, within the target age range. Therefore, the rankings canbe used to create models for a general population of cell phoneusers, in the given age range, that respond to Internet surveys.

6. The KE-QT1 approach

The KE-QT1 approach creates Kansei value user need descrip-tions, creates a KE-QT1 statistical model, creates Kansei value de-sign descriptions, and matches Kansei value design descriptionsto Kansei value user need descriptions (Table 2).

6.1. Create user need descriptions

The KE-QT1 approach creates emotional (Kansei value) userneed descriptions. In the case study, the KE-QT1 user need descrip-tion, for any user, is a set that contains six Kansei values. The sixKansei values correspond to the six Kansei words, in Table 3.

Designers created KE-QT1 user need descriptions, for all 79 casestudy users. The case study KE-QT1 user need description for User39 is (2, 3, 4, 3, 3, 3).

6.2. Create a model

The KE-QT1 approach creates a KE-QT1 statistical model. TheKE-QT1 model shows the impact of each sample design elementon the sample design rankings. The KE-QT1 approach also usesthe model to calculate Kansei values for cell phone designs. Inthe case study, designers used QT1 analysis to create a KE-QT1 sta-tistical model from the Table 7 sample design rankings (‘design byKansei word matrix’).

Table 8 QT1 analysis results show the impact of each sample de-sign element on sample design ‘elegant’ rankings. Multiple correla-tion coefficient (R) and coefficient of determination (R2) show thatdesign elements had a significant impact on ‘elegant’ rankings,R = 0.9954, R2 = 0.9908. Constant value (K) shows that users rankedthe sample designs slightly ‘ordinary’, K = 4.2440.

Partial correlation coefficients (PCC), show that ‘body shape’ hadthe greatest impact on ‘elegant’ rankings, PCC = 0.9943. Associateddesign element weights show that most subjects ranked cellphones with ‘straight’ body shapes more ‘elegant’ and cell phoneswith curved (‘round’ or ‘arc’) body shapes more ‘ordinary’ (bold).

Table 7Kansei values for samples.

Sample ‘elegant’ ‘simple’ ‘high tech’ ‘luxurious’ ‘beautiful’ ‘unique’

Sample 1 3.674 2.174 3.700 5.217 3.864 5.174Sample 2 4.609 3.325 4.675 5.043 5.000 4.826Sample 3 5.043 4.370 4.450 4.587 5.543 3.870Sample 4 3.978 3.674 4.000 4.565 4.065 4.478Sample 5 4.804 4.435 4.100 4.304 5.130 3.196Sample 6 3.783 3.326 3.275 4.000 3.652 3.891Sample 7 3.217 3.870 3.000 3.261 3.152 3.435Sample 8 4.326 4.717 3.575 3.609 4.761 3.022Sample 9 4.239 4.739 3.450 3.975 4.935 2.761Sample 10 5.456 3.435 4.700 5.130 5.609 4.043Sample 11 3.391 2.891 3.500 4.326 3.630 4.522Sample 12 5.022 3.696 4.600 4.761 5.304 4.261Sample 13 5.000 4.739 3.925 4.065 4.935 3.478Sample 14 3.109 3.438 3.325 3.978 3.152 4.152Sample 15 4.500 3.761 4.200 4.500 4.522 3.891Sample 16 4.480 4.217 3.725 4.152 4.500 3.717Sample 17 4.717 4.739 3.600 3.696 5.000 3.130Sample 18 3.043 2.913 3.200 3.935 2.826 4.457

Samples (p-value = 0.0002), Kansei values (p-value = 0.0048).

Table 8KE-QT1 analysis for ‘elegant’.

Element Level Weight Graph PCC

‘vertical’ -0.289 ‘phone type’ ‘folding’ 0.048 0.954

‘sliding’ 0.240

‘1:1’ -0.057

‘screen shape’ ‘25:16’ 0.085 0.665

‘4:3’ -0.028

‘straight’ 0.161

‘top shape’ ‘round’ -0.039 0.866

‘arc’ -0.122


‘body shape’ ‘round’ -0.242 0.994

‘arc’ -0.633

‘straight’ -0.176

‘bottom shape’ ‘round’ 0.172 0.901

‘arc’ 0.005


‘number keys’ ‘round’ -0.136 0.833

‘arc’ 0.114

‘5 keys’ -0.057

‘function keys’ ‘7 keys’ -0.006 0.582

‘8 keys' 0.063

K = 4.244, R = 0.995, R2 = 0.991.


Table 9 shows the complete KE-QT1 model for all Kansei wordpairs. Each matrix element corresponds to a QT1 design elementweight. The QT1 design element weight for the second word ofeach pair is the same as that for the first word, with opposite sign.The KE-QT1 model only includes one design element weight foreach Kansei word pair. The KE-QT1 approach uses the model to cal-culate Kansei values for new designs.

When used to calculate Kansei values for the eighteen sampledesigns, the KE-QT1 model matched the Kansei values in Table 7with 87.4% accuracy. The results show that the KE-QT1 approachcan create design descriptions with 87.4% model accuracy (MA).As a result, on average, the KE-QT1 approach can match designsto user needs with up to 87.4% matching accuracy (UCDA).

6.3. Create design descriptions

The KE-QT1 approach creates emotional (Kansei value) designdescriptions. The KE-QT1 approach chooses design elements,

create designs, calculates Kansei values for the designs, createsKansei value design descriptions, uses statistical correlation tomatch Kansei value design descriptions to Kansei value user needdescriptions, and chooses the design, for each user, that matchesuser needs most accurately.

In the case study, the KE-QT1 design description, for any cellphone, contains six Kansei values. To create design descriptions,designers analyzed the twelve Table 4 cell phones, identifiedTable 5 design elements, used the Table 9 KE-QT1 model to calcu-late KE-QT1 Kansei values, and used the KE-QT1 Kansei values tocreate Kansei value design descriptions.

To choose designs for each users, designers used correlation tomatch the Kansei value design descriptions to Kansei value userneed descriptions. Table 10 shows KE-QT1 Kansei values for oneuser (User 39) and all 12 Table 4 cell phones.

6.4. KE-QT1 matching results

Table 11 shows user choice and KE-QT1 choice for one user(User 39), user choices for all users, KE-QT1 choices for all users,and KE-QT1 CMA, UCDA, and CDA. For one user, User 39 chosePhone 5 (bold). The KE-QT1 approach chose Phone 11 (bold).

User 39 optimized CDA. User 39 chose the phone that matcheduser concept (user needs and Phone 5 design elements) most accu-rately. The KE-QT1 approach optimized correlation to optimizeUCDA. The KE-QT1 approach tried to choose the phone thatmatched user needs most accurately. The KE-QT1 approach didnot optimize UCDA or CDA accurately, (p-values = 0.0000, 0.0000,0.0865, 0.6191).

For all users, users chose Phone 3 most often (bold). The KE-QT1approach chose Phone 7 most often (bold in Table 11). Users chosephones with higher CDA and lower UCDA (bold). Users chosephones that matched user concepts more accurately. The KE-QT1approach chose phones with higher UCDA and lower CDA (bold).The KE-QT1 approach chose phones that matched user needs moreaccurately, (p-values = 0.0000) (bold).

The KE-QT1 approach matched user choices, needs, and con-cepts with (7.6% CMA, 83.5% UCDA, 64.1% CDA). The KE-QT1 ap-proach chose phones with higher UCDA than user choices. TheKE-QT1 chose phones with lower CMA and CDA than user choices,(p-values = 0.0000) (bold). The goal of this study is to improvematching accuracy, compared to the KE-QT1 approach.

7. Latent Semantic Analysis (LSA)

The LSA document indexing approach creates a ‘term by docu-ment’ matrix, converts the matrix into a LSA semantic space model,projects keyword term vectors into the semantic space, andmatches projected keyword term vectors to projected documentvectors, in the LSA semantic space. The approach improves docu-ment recall and precision over simple term matching document re-trieval systems [12].

7.1. Create ‘term by document’ matrix

Table 12 shows an example ‘term by document’ matrix for twodocuments related to information retrieval (Documents 1 and 3),one document related to information theory (Document 2), andeight keyword terms. Bold rows indicates keyword terms (‘com-puter’ and ‘information’) that are contained in one or moredocuments.

For the keyword search, ‘IDF in computer-based informationlook-up’, a term matching approach would miss Document 1, a re-call or synonymy problem, and return Document 2, a precision orpolysemy problem [12]. The LSA approach reduces synonymy

Table 9KE-QT1 model.

Element Level ‘elegant’ ‘simple’ ‘high tech’ ‘luxurious’ ‘beautiful’ ‘unique’

‘vertical’ �0.289 0.488 �0.438 �0.561 �0.404 �0.543‘phone type’ ‘folding’ 0.048 �0.092 0.029 0.048 0.178 0.058

‘sliding’ 0.240 �0.396 0.408 0.512 0.226 0.486

‘1:1’ �0.057 0.118 �0.008 �0.115 0.067 �0.149‘screen shape’ ‘25:16’ 0.085 �0.121 0.038 0.124 �0.024 0.098

‘4:3’ �0.028 0.003 �0.029 �0.009 �0.043 0.051

‘straight’ 0.161 0.140 0.092 �0.096 0.085 �0.087‘top shape’ ‘round’ �0.039 �0.026 �0.083 0.055 0.030 �0.036

‘arc’ �0.122 �0.114 �0.008 0.041 �0.115 0.123

‘straight’ 0.874 0.701 0.500 0.164 1.042 �0.366‘body shape’ ‘round’ �0.242 �0.095 �0.296 �0.155 �0.271 �0.127

‘arc’ �0.633 �0.606 �0.204 �0.009 �0.771 0.493

‘straight’ �0.176 �0.095 0.054 0.059 �0.165 0.080‘bottom shape’ ‘round’ 0.172 0.216 �0.058 �0.154 0.157 �0.214

‘arc’ 0.005 �0.121 0.004 0.095 0.008 0.134

‘straight’ 0.023 0.020 0.017 �0.114 �0.060 0.029‘number keys’ ‘round’ �0.136 0.061 �0.075 0.012 �0.068 0.036

‘arc’ 0.114 �0.081 0.058 0.102 0.128 �0.065

‘5 keys’ �0.057 0.231 �0.079 �0.166 �0.031 �0.058‘function keys’ ‘7 keys’ �0.006 �0.070 0.075 0.023 �0.021 0.040

‘8 keys’ 0.063 �0.161 0.004 0.142 0.052 0.018

KE-QT1 model accuracy = 87.4%.

Table 10KE-QT1 Kansei values.

‘elegant’ ‘simple’ ‘high tech’ ‘luxurious’ ‘beautiful’ ‘unique’

User 2.000 3.000 4.000 3.000 3.000 3.000

Phone 1 5.283 3.887 4.508 4.895 5.539 3.971Phone 2 3.692 2.492 3.879 4.708 3.581 4.989Phone 3 3.685 2.616 3.563 4.818 3.986 4.703Phone 4 5.116 3.551 4.571 5.145 5.390 4.319Phone 5 3.236 3.320 3.029 4.076 3.386 4.054Phone 6 4.776 3.931 4.288 4.417 4.762 3.688Phone 7 3.576 3.047 3.567 4.084 3.304 4.087Phone 8 4.504 4.257 3.979 4.243 4.447 3.891Phone 9 4.674 5.188 3.529 3.528 5.009 2.645Phone 10 3.156 3.500 2.717 3.745 3.357 3.674Phone 11 2.946 3.178 3.092 3.862 3.082 3.967Phone 12 3.218 3.580 2.979 3.543 3.207 3.601


and polysemy problems by projecting terms, documents, and key-words into a semantic space that models term-document relation-ships better than term matching.

7.2. Create a LSA semantic space model

The LSA approach uses singular value decomposition (SVD) toconvert the ‘term by document’ matrix into a semantic space mod-el. The approach uses dimension reduction to remove noise andimprove results. The approach uses a cosine measure to match vec-tors in the semantic space. The approach can project new docu-ment vectors or new term vectors into the semantic space, toupdate the semantic space model.

For an m � n ‘term by document’ matrix, M0 that records thenumber of times m terms occur in n documents, the SVD of M0 is

M0 ¼ T0S0D00 ð1Þ

T0 is an m � n matrix, S0 is an n � n diagonal matrix, and D0 is ann � n matrix. Rows in T0 represent m term vectors in the n-dimen-sional semantic space. Rows in D0 represent n document vectors inthe n-dimensional semantic space. The diagonal values in S0 repre-

sent scaling factors for the vectors, with respect to the ndimensions.

7.3. Reduce dimension

By convention, the diagonal values in S0 are all positive and or-dered in decreasing magnitude. Dimension reduction consists ofdeleting rows and columns from S0 below a threshold value anddeleting the corresponding columns from both T0 and D0, to createreduced S, T and D matrices and a reduced model M

M0 � M ¼ TSD0 ð2Þ

Reducing dimensions reduces noise effects. However, if thenumber of dimensions is too small, important information maybe lost [13]. For any given problem, the amount of information inthe reduced model is

Information ¼

Pk

i¼1si

Pn

i¼1si

� 100% ð3Þ

si is the ith diagonal value in S0, n is the dimension of S0, and k is thedimension of S. Most LSA studies choose k for best performance[12].

7.4. Match document vectors

For a given LSA problem, the reduced ‘term by document’ ma-trix M is a LSA semantic space, which contains m term vectorsand n document vectors, with the same dimension as S. Matchingdocuments to documents consists of computing the cosine of theangle between scaled document vectors in the semantic space

DiS � DjSDiSj j � DjS

��

�� ð4Þ

Di and Dj are the ith and jth row vectors in the D matrix.A keyword search consists of creating an m-element ‘term by

document’ vector Mv, projecting Mv into the semantic space, andcomparing the scaled semantic space vector Dv to scaled document

Table 11KE-QT1matching results.

User choice and KE-QT1 choice, one userPhone 11 2 7 3 5 4 12 6 9 1 10 8

Corr 0.105 0.066 �0.007 �0.048 �0.148 �0.253 �0.294 �0.352 �0.358 �0.360 �0.366 �0.678CMA 0.0 0.0 0.0 0.0 100.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0UCDA 88.1 57.4 80.9 54.9 80.9 13.8 88.2 16.5 36.6 11.9 83.7 7.5CDA 62.6 43.3 52.9 80.2 99.4 32.7 44.0 42.4 42.7 32.6 80.9 51.9

User choices, all usersPhone 3 5 10 12 11 4 6 1 2 8 7 9

Percent 24.1 22.8 16.5 11.39 8.9 3.8 3.8 2.5 2.5 2.5 1.3 0.0CMA 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0UCDA 69.6 79.5 89.1 83.6 83.8 65.0 38.6 20.7 74.3 46.3 72.5CDA 98.1 98.9 98.9 98.6 98.8 98.5 96.6 94.5 98.5 98.2 99.6

KE-QT1 choices, all usersPhone 7 11 12 5 3 2 8 9 10 6 1 4

Percent 30.4 15.2 12.7 10.1 8.9 5.1 5.1 5.1 5.1 2.5 0.0 0.0CMA 0.0 0.0 20.0 12.5 14.3 0.0 50.0 0.0 0.0 0.0UCDA 86.9 91.5 89.7 92.3 80.6 75.0 40.1 69.8 90.1 56.1CDA 57.4 60.6 64.6 81.0 67.5 57.7 77.2 57.3 80.8 51.9

KE-QT1 accuracyApproach Users KE-QT1 p-value

CMA 100.0 7.6 0.0000UCDA 74.9 83.5 0.0000CDA 98.4 64.1 0.0000

Table 12LSA ‘term by document’ matrix [12].

Term Document 1 Document 2 Document 3

‘access’ 1 0 0‘computer’ 0 1 1‘database’ 1 0 0‘document’ 1 0 0‘indexing’ 1 0 0‘information’ 0 1 1‘retrieval’ 1 0 1‘theory’ 0 1 0

Search: ‘IDF in computer-based information look-up’.

Table 13LSA keyword search.

Mv = T S Dv‘

Table 14KE-LSA D0 matrix.

�0.137 0.669 �0.080 �0.045 0.320 0.1800.049 �0.091 0.212 0.083 �0.477 �0.3320.089 �0.579 �0.132 �0.038 0.156 0.1520.034 0.125 0.001 0.200 �0.019 �0.571�0.006 �0.124 0.083 �0.224 0.157 0.396�0.028 �0.001 �0.084 0.025 �0.137 0.175

0.104 0.065 �0.195 �0.424 0.040 �0.117�0.010 0.004 0.285 0.298 �0.036 0.131�0.094 �0.070 �0.089 0.126 �0.003 �0.015

0.750 0.033 �0.176 0.111 0.084 0.038�0.189 0.166 0.426 �0.029 �0.020 0.018�0.560 �0.199 �0.250 �0.082 �0.064 �0.056�0.105 �0.059 �0.201 0.246 0.467 �0.245

0.131 0.176 0.138 �0.281 �0.268 0.057�0.025 �0.116 0.063 0.035 �0.199 0.188�0.015 0.033 �0.225 �0.403 �0.177 �0.081�0.056 0.037 �0.124 0.492 �0.075 0.243

0.071 �0.069 0.349 �0.089 0.252 �0.1610.004 0.171 �0.323 0.162 �0.324 0.149�0.012 �0.057 �0.070 �0.113 0.168 �0.251

0.009 �0.114 0.393 �0.049 0.156 0.102

Table 15KE–LSA S0 matrix.

2.210 0.000 0.000 0.000 0.000 0.0000.000 1.629 0.000 0.000 0.000 0.0000.000 0.000 0.334 0.000 0.000 0.0000.000 0.000 0.000 0.220 0.000 0.0000.000 0.000 0.000 0.000 0.185 0.0000.000 0.000 0.000 0.000 0.000 0.159


vectors in the semantic space (Table 13). Mv is the column vector ofterms in the keyword search. Creating Dv consists of calculating

Dv ¼ M0vTS�1 ð5Þ

Completing the keyword search consists of computing the co-sine of the angle between the scaled Dv vector and scaled docu-ment vectors in the semantic space

DvS � DjSDvSj j � DjS

��

�� ð6Þ

7.5. Update the model

The LSA ‘term by document matrix’ M0 model generally con-tains a large number of terms and documents. Therefore, the LSAapproach does not generally re-compute the SVD of M0 every time

the approach adds new terms or documents to the matrix. Instead,the approach projects new terms or documents into the existingsemantic space.

Projecting a new document into the existing semantic spaceconsists of creating an m-element document vector Md, projectingMd into the semantic space, and adding the scaled semantic spacevector Dd to the D matrix [14]. Md is the column vector of terms inthe document. Creating Dd consists of calculating

Dd ¼ M0dTS�1 ð7Þ

Table 16KE-LSA W0 matrix.

0.544 �0.129 0.105 �0.687 0.001 0.4520.380 0.532 �0.553 0.359 0.027 0.3700.293 �0.345 �0.544 �0.131 0.455 �0.5240.110 �0.510 0.231 0.508 0.471 0.4400.644 �0.145 0.274 0.351 �0.496 �0.347�0.217 �0.547 �0.508 0.027 �0.569 0.266

Table 17KE-LSA Kansei vectors.

User [2.000, 1.000, 0.000, 1.000, 1.000, 1.000]

Phone 1 [�1.283, 0.113, �0.508, �0.895, �1.539, 0.029]Phone 2 [0.308, 1.508, 0.121, �0.708, 0.419, �0.989]Phone 3 [0.315, 1.384, 0.437, �0.818, 0.014, �0.703]Phone 4 [�1.116, 0.449, �0.571, �1.145, �1.390, �0.319]Phone 5 [0.764, 0.680, 0.971, �0.076, 0.614, �0.054]Phone 6 [�0.776, 0.069, �0.287, �0.417, �0.762, 0.312]Phone 7 [0.424, 0.953, 0.433, �0.084, 0.696, �0.087]Phone 8 [�0.504, �0.257, 0.021, �0.243, �0.447, 0.109]Phone 9 [�0.674, �1.188, 0.471, 0.472, �1.009, 1.355]Phone 10 [0.844, 0.500, 1.283, 0.255, 0.643, 0.326]Phone 11 [1.054, 0.822, 0.908, 0.138, 0.918, 0.033]Phone 12 [0.782, 0.420, 1.021, 0.457, 0.793, 0.399]

Table 18KE-LSA projected Kansei vectors.

User [2.004, �0.929, �0.345, �0.130, �0.566, 1.632]

Phone 1 [�1.900, 1.065, �0.564, �0.005, 0.096, �0.124]Phone 2 [1.182, 1.563, �0.414, 0.075, 0.117, �0.086]Phone 3 [0.897, 1.345, �0.799, �0.206, 0.244, �0.126]Phone 4 [�1.556, 1.541, �0.539, �0.075, 0.083, �0.145]Phone 5 [1.358, �0.093, �0.646, �0.232, 0.151, �0.173]Phone 6 [�1.084, 0.389, �0.427, 0.125, �0.126, �0.011]Phone 7 [1.177, 0.292, �0.503, 0.193, �0.112, 0.016]Phone 8 [�0.704, 0.051, �0.156, �0.026, 0.047, �0.257]Phone 9 [�1.572, �1.544, �0.525, �0.103, 0.134, �0.073]Phone 10 [1.396, �0.688, �0.817, �0.204, 0.213, �0.131]Phone 11 [1.751, �0.234, �0.571, �0.155, 0.027, 0.055]Phone 12 [1.359, �0.796, �0.585, 0.001, 0.071, 0.005]

Table 19KE-LSA matching results.

User choice and KE-LSA choice, one userPhone 12 11 10 5 7 2

Cos 0.974 0.945 0.945 0.897 0.753 0.234CMA 0.0 0.0 0.0 100.0 0.0 0.0UCDA 88.2 88.1 83.7 80.9 80.9 57.4CDA 44.0 62.6 80.9 99.4 52.9 43.3

User choices, all usersPhone 3 5 10 12 11 4

Percent 24.1 22.8 16.5 11.4 8.9 3.8CMA 100.0 100.0 100.0 100.0 100.0 100.0UCDA 69.6 79.5 89.1 83.6 83.8 65.0CDA 98.1 98.9 98.9 98.6 98.8 98.5

KE-LSA choices, all usersPhone 12 7 11 10 2 5


KE-LSA accuracyApproach Users KE-QT1

CMA 100.0 7.6UCDA 74.9 83.5CDA 98.4 64.1


The approach adds new terms to T in a similar manner.

8. The KE-LSA approach

The KE-LSA approach creates Kansei vector user need descrip-tions, creates a KE-LSA semantic space model, creates Kansei vectordesign descriptions, projects Kansei vectors into the KE-LSAsemantic space, and matches projected Kansei vectors, in the KE-LSA semantic space (Table 2).


The KE-LSA approach creates emotional (Kansei vector) userneed descriptions. In the case study, the KE-LSA user need descrip-tion, for any user, is a six-element zero-mean Kansei vector. Eachvector element is derived from a Section 5.1 user need

ui ¼ �1� user needi � survey meanð Þ ð8Þ

Designers created KE-LSA user need descriptions, for all 79 casestudy users. The case study KE-LSA user need description for User39 is [2,1,0,1,1,1].

8.2. Create a model

The KE-LSA approach uses SVD to create a KE-LSA semanticspace model from a KE-QT1 model. The KE-LSA approach usesthe KE-LSA model to project Kansei vectors into the KE-LSA seman-tic space. In the case study, designers used SVD to create a KE-LSAsemantic space model from the Table 9 KE-QT1 model

M0 ¼ D0S0W 00 ð9Þ

D0 is a 21 � 6 matrix, S0 is a 6 � 6 diagonal matrix, and W0 is a 6 � 6matrix (Tables 14–16). Rows in D0 and W0 represent projected de-sign and Kansei word vectors in the six-dimensional semanticspace. The diagonal values in S0 represent scaling factors for thevectors, with respect to the six dimensions. The diagonal values inS0 are all positive and ordered in decreasing magnitude.

3 9 4 8 6 1

0.209 �0.300 �0.861 �0.865 �0.869 �0.9130.0 0.0 0.0 0.0 0.0 0.0

54.9 36.6 13.8 7.5 16.5 11.980.2 42.7 32.7 51.9 42.4 32.6

6 1 2 8 7 9

3.8 2.5 2.5 2.5 1.3 0.0100.0 100.0 100.0 100.0 100.038.6 20.7 74.3 46.3 72.596.6 94.5 98.5 98.2 99.6

3 1 4 9 6.0 8

6.3 1.3 1.3 1.3 0.0 0.00.0 0.0 100.0 0.0

84.4 82.9 76.0 84.262.7 43.8 99.5 62.3

KE-LSA p-value p-value

12.7 0.0000 0.294690.4 0.0000 0.001063.9 0.0000 0.9226

Table 20KE-LSA matching accuracy, k = 2–6.

Approach Dimension Information CMA UCDA CDA

KE-QT1 k = 6 100.0 7.6 83.5 64.1

KE-LSA k = 2 81.0 8.9 89.5 64.8KE-LSA k = 3 88.1 10.1 89.5 63.7KE-LSA k = 4 92.7 12.7 90.4 63.9KE-LSA k = 5 96.6 12.7 89.5 65.1KE-LSA k = 6 100.0 10.1 89.5 63.3

Table 21LSE design descriptions for samples.

1 2 3 4 5 6 7 8 9 10

0.33 �0.61 �1.04 0.02 �0.80 0.22 0.78 �0.33 �0.24 �11.83 0.68 �0.37 0.33 �0.43 0.67 0.13 �0.72 �0.74 00.30 �0.68 �0.45 0.00 �0.10 0.73 1.00 0.43 0.55 �0�1.22 �1.04 �0.59 �0.57 �0.30 0.00 0.74 0.39 0.02 �1

0.14 �1.00 �1.54 �0.07 �1.13 0.35 0.85 �0.76 �0.93 �1�1.17 �0.83 0.13 �0.48 0.80 0.11 0.57 0.98 1.24 �0

1 1 1 0 0 0 0 0 0 10 0 0 1 1 1 0 0 0 00 0 0 0 0 0 1 1 1 01 0 0 1 0 0 1 0 0 10 1 0 0 1 0 0 1 0 00 0 1 0 0 1 0 0 1 01 0 0 1 0 0 0 0 1 00 1 0 0 1 0 1 0 0 00 0 1 0 0 1 0 1 0 11 0 0 0 0 1 1 0 0 00 1 0 1 0 0 0 1 0 00 0 1 0 1 0 0 0 1 11 0 0 0 0 1 0 1 0 00 1 0 1 0 0 0 0 1 10 0 1 0 1 0 1 0 0 01 0 0 0 1 0 0 0 1 00 1 0 0 0 1 1 0 0 10 0 1 1 0 0 0 1 0 01 0 0 0 1 0 0 1 0 10 1 0 0 0 1 0 0 1 00 0 1 1 0 0 1 0 0 0

Table 22LSE E0 matrix.

0.06 �0.18 0.04 0.01 0.03 0.00 �0.01 0.00 0.00�0.04 �0.14 �0.14 �0.03 �0.01 0.00 �0.01 0.00 0.00�0.03 �0.10 0.09 �0.03 0.00 0.01 0.01 0.00 0.00

0.07 �0.03 0.13 0.01 �0.02 0.01 �0.01 0.00 0.000.10 �0.22 0.04 0.01 �0.01 �0.02 0.01 0.00 0.00�0.02 0.08 0.14 �0.03 0.00 �0.02 �0.01 0.00 0.00

0.20 0.13 �0.66 �0.08 0.05 0.32 �0.18 �0.04 �0.03 �0.22 �0.03 0.08 0.21 �0.08 �0.48 0.33 �0.03 0.25 �0.23 �0.06 0.55 �0.13 0.04 0.16 �0.15 0.07 �0.220.21 �0.02 �0.13 0.00 �0.20 �0.02 0.57 0.16 �0.19 �0.22 0.02 0.11 0.08 0.22 0.16 �0.39 0.03 0.440.21 0.04 �0.01 �0.09 �0.02 �0.14 �0.17 �0.19 �0.26 �0.22 �0.09 �0.07 �0.20 0.42 0.04 0.12 0.11 �0.04 �0.22 0.02 �0.01 0.28 �0.30 �0.04 �0.13 0.13 �0.10 �0.21 0.10 0.06 �0.09 �0.13 0.00 0.02 �0.24 0.140.24 �0.69 �0.03 �0.18 �0.11 0.08 �0.04 �0.03 0.02 �0.21 0.19 �0.17 0.42 0.03 �0.02 �0.02 0.12 0.010.19 0.54 0.17 �0.25 0.08 �0.06 0.06 �0.08 �0.03 �0.21 0.11 0.05 �0.20 �0.25 0.47 0.25 0.23 0.090.22 �0.11 �0.17 0.13 0.28 �0.27 �0.06 �0.32 �0.220.22 0.04 0.10 0.06 �0.03 �0.20 �0.19 0.08 0.14 �0.21 0.03 �0.04 �0.23 0.40 �0.17 0.08 0.31 �0.180.21 0.07 �0.04 �0.13 �0.49 �0.07 �0.24 �0.24 �0.19 �0.22 �0.05 0.06 0.35 0.09 0.25 0.16 �0.07 0.370.21 0.01 �0.17 �0.33 �0.16 �0.32 �0.15 0.33 0.330.22 0.03 0.05 �0.07 0.11 0.17 0.25 �0.58 0.07 �0.22 0.01 0.10 0.39 0.05 0.16 �0.10 0.24 �0.40 �


When used to calculate and project Kansei values for sampledesigns into and out of the KE-LSA semantic space, the KE-QT1and KE-LSA models matched the Kansei values in Table 7 with87.4% accuracy. The KE-LSA approach can create design descrip-tions with 87.4% model accuracy (MA). On average, the KE-LSA ap-proach can match designs to user needs with up to 87.4% matchingaccuracy (UCDA).

11 12 13 14 15 16 17 18

.46 0.61 �1.02 �1.00 0.89 �0.50 �0.48 �0.72 0.96

.57 1.11 0.30 �0.74 0.56 0.24 �0.22 �0.74 1.09

.70 0.50 �0.60 0.08 0.68 �0.20 0.28 0.40 0.80

.13 �0.33 �0.76 �0.07 0.02 �0.50 �0.15 0.30 0.07

.61 0.37 �1.30 �0.93 0.85 �0.52 �0.50 �1.00 1.17

.04 �0.52 �0.26 0.52 �0.15 0.11 0.28 0.87 �0.461 1 0 0 0 0 0 00 0 1 1 1 0 0 00 0 0 0 0 1 1 10 0 1 0 0 1 0 01 0 0 1 0 0 1 00 1 0 0 1 0 0 11 0 0 0 1 0 1 00 1 1 0 0 0 0 10 0 0 1 0 1 0 01 0 0 1 0 0 0 10 1 0 0 1 1 0 00 0 1 0 0 0 1 00 1 1 0 0 0 1 00 0 0 1 0 0 0 11 0 0 0 1 1 0 00 1 0 1 0 1 0 00 0 0 0 1 0 1 01 0 1 0 0 0 0 10 0 0 0 1 0 0 11 0 1 0 0 1 0 00 1 0 1 0 0 1 0

0.00 0.00 0.00 0.00 0.00 0.00 0.09 0.75 �0.250.00 0.00 0.00 0.00 0.00 0.00 �0.45 �0.48 �0.550.00 0.00 0.00 0.00 0.00 0.00 0.15 0.16 �0.260.00 0.00 0.00 0.00 0.00 0.00 0.36 �0.28 0.390.00 0.00 0.00 0.00 0.00 0.00 �0.16 �0.06 0.530.00 0.00 0.00 0.00 0.00 0.00 0.74 �0.29 �0.350.15 �0.02 �0.08 0.04 �0.11 0.10 0.14 0.02 0.030.01 0.19 0.39 0.13 �0.03 �0.03 �0.02 �0.02 �0.040.17 �0.17 �0.31 �0.17 0.14 �0.07 �0.13 �0.03 �0.060.06 �0.31 �0.28 0.03 �0.14 �0.15 0.03 0.00 �0.020.11 0.03 0.14 0.09 �0.26 �0.32 �0.03 �0.02 �0.030.05 0.28 0.14 �0.13 0.40 0.47 �0.01 0.00 �0.020.17 �0.15 0.20 0.20 0.46 �0.26 0.02 �0.03 0.000.02 0.54 �0.36 0.05 �0.10 �0.13 0.00 0.00 �0.040.19 �0.39 0.16 �0.25 �0.35 0.39 �0.03 0.00 �0.020.09 0.07 0.04 0.07 �0.15 0.09 0.09 �0.07 �0.020.17 �0.14 0.07 �0.50 0.26 �0.17 0.02 0.01 �0.030.08 0.07 �0.12 0.43 �0.12 0.08 �0.11 0.04 �0.010.23 0.25 0.29 �0.04 0.09 0.01 �0.03 0.01 �0.020.46 �0.06 �0.18 0.21 �0.06 �0.05 0.05 �0.03 �0.010.69 �0.19 �0.11 �0.17 �0.03 0.04 �0.03 �0.01 �0.030.01 0.24 0.04 �0.33 �0.34 0.03 0.00 �0.01 0.000.01 �0.18 0.23 0.11 0.10 �0.37 0.00 0.02 �0.020.00 �0.06 �0.27 0.22 0.24 0.34 �0.01 �0.03 �0.040.24 �0.08 �0.22 0.02 0.20 0.08 0.03 0.01 0.000.17 0.22 �0.09 �0.27 �0.02 �0.24 �0.02 �0.01 �0.020.07 �0.14 0.31 0.25 �0.18 0.16 �0.02 �0.02 �0.05

Table 23LSE S0 matrix.

19.64 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 15.65 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 15.23 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 14.72 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 0.00 14.71 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 14.70 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 14.70 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 14.70 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 14.70 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 14.70 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 14.70 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 14.70 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 14.70 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 14.70 0.00 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 14.70 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.20 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.35 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.22

Table 24LSE D0 matrix.

�0.24 �0.28 �0.41 �0.49 0.06 0.16 0.27 0.44 0.00 0.00 0.00 0.00 0.00 0.00 �0.04 �0.13 0.38 �0.02�0.25 0.09 �0.35 0.27 �0.04 0.10 �0.32 �0.37 �0.01 0.16 0.16 �0.11 �0.11 �0.08 �0.49 �0.14 0.38 �0.02�0.25 0.30 �0.04 0.12 0.03 0.13 �0.17 �0.12 �0.03 �0.35 �0.18 0.02 0.16 �0.06 0.65 �0.15 0.38 �0.02�0.23 �0.10 �0.10 0.54 0.24 �0.14 0.41 0.09 �0.09 0.13 �0.28 0.10 0.22 0.23 �0.06 �0.04 �0.04 �0.40�0.24 0.22 0.13 �0.06 0.06 �0.46 �0.16 0.31 0.35 �0.18 0.32 �0.10 0.09 �0.28 �0.10 �0.05 �0.04 �0.40�0.22 �0.22 0.09 �0.22 �0.40 0.01 0.16 �0.44 0.05 0.04 0.18 0.48 �0.16 0.02 0.14 �0.04 �0.04 �0.40�0.21 �0.33 0.27 0.13 �0.43 0.03 �0.12 0.17 �0.39 �0.31 �0.16 �0.20 0.07 �0.15 �0.18 0.38 0.14 �0.06�0.23 0.09 0.24 0.05 �0.06 0.28 �0.12 0.19 0.48 0.45 �0.23 �0.06 �0.26 0.13 0.10 0.37 0.14 �0.05�0.24 0.11 0.24 �0.33 0.54 �0.11 0.05 �0.27 �0.36 0.07 0.18 �0.13 �0.03 0.19 �0.01 0.37 0.14 �0.05�0.26 0.26 �0.37 �0.30 �0.23 �0.18 0.01 �0.18 �0.08 0.24 �0.40 �0.20 0.25 �0.20 0.03 0.15 �0.38 0.02�0.22 �0.32 �0.15 �0.01 0.30 0.34 �0.11 �0.20 0.39 �0.47 �0.04 �0.07 0.08 0.05 �0.10 0.15 �0.38 0.02�0.25 0.18 �0.28 0.21 �0.02 0.23 �0.11 0.33 �0.35 0.05 0.41 0.17 �0.27 0.01 0.19 0.15 �0.38 0.02�0.24 0.21 0.14 0.14 �0.22 0.12 0.61 �0.09 0.18 �0.05 0.37 �0.18 0.22 �0.03 �0.05 0.04 0.04 0.40�0.21 �0.35 0.08 0.14 0.26 �0.22 �0.07 �0.02 0.02 0.24 �0.03 0.37 0.07 �0.56 0.11 0.05 0.04 0.40�0.24 0.04 �0.09 �0.01 �0.14 �0.49 �0.13 0.07 0.10 �0.21 �0.11 0.29 �0.13 0.56 �0.10 0.05 0.04 0.40�0.24 0.07 0.20 �0.01 0.09 �0.04 0.23 �0.03 �0.09 �0.15 �0.30 �0.21 �0.68 �0.20 �0.07 �0.38 �0.14 0.05�0.24 0.21 0.39 �0.17 0.03 0.34 �0.19 0.15 �0.14 0.07 �0.12 0.30 0.36 0.09 �0.32 �0.39 �0.14 0.05�0.21 �0.42 0.11 0.03 �0.08 �0.11 �0.22 �0.03 �0.03 0.29 0.21 �0.47 0.11 0.28 0.30 �0.38 �0.14 0.05

Table 25LSE user need and design descriptions.

User 1 2 3 4 5 6 7 8 9 10 11 12

2.00 0 0 0 0 0 0 0 0 0 0 0 01.00 0 0 0 0 0 0 0 0 0 0 0 00.00 0 0 0 0 0 0 0 0 0 0 0 01.00 0 0 0 0 0 0 0 0 0 0 0 01.00 0 0 0 0 0 0 0 0 0 0 0 01.00 0 0 0 0 0 0 0 0 0 0 0 00 1 1 1 1 0 0 0 0 0 0 0 00 0 0 0 0 1 1 1 1 0 0 0 00 0 0 0 0 0 0 0 0 1 1 1 10 1 1 0 1 0 0 0 1 0 0 0 00 0 0 0 0 1 0 0 0 0 0 1 00 0 0 1 0 0 1 1 0 1 1 0 10 0 0 1 0 1 0 0 0 0 1 0 00 0 1 0 0 0 0 1 0 0 0 1 00 1 0 0 1 0 1 0 1 1 0 0 10 0 1 1 0 1 0 1 0 0 1 1 10 0 0 0 0 0 1 0 1 0 0 0 00 1 0 0 1 0 0 0 0 1 0 0 00 0 0 1 0 1 0 0 0 0 1 0 00 0 0 0 1 0 0 0 0 0 0 1 00 1 1 0 0 0 1 1 1 1 0 0 10 0 0 1 0 1 0 0 0 0 1 1 00 0 0 0 0 0 1 1 0 0 0 0 00 1 1 0 1 0 0 0 1 1 0 0 10 1 1 0 1 0 1 1 0 0 0 1 10 0 0 1 0 1 0 0 1 1 1 0 00 0 0 0 0 0 0 0 0 0 0 0 0



The KE-LSA approach creates emotional (Kansei vector) designdescriptions. The KE-LSA approach chooses design elements, cre-ates designs, calculates Kansei values for the designs, creates Kan-sei vectors for the designs, projects Kansei vector user needs andKansei vector designs into the KE-LSA semantic space, matchesprojected vectors, and chooses the best design for each user.

In the case study, designers projected Kansei vectors into theKE-LSA semantic space by multiplying each Kansei vector by trans-formation matrices, according to Eq. (9). Designers matched pro-jected vectors, in the KE-LSA semantic space, by calculating thecosine of the angle between the projected vectors

DuS � DdSjDuSj � jDdSj ð10Þ

Du is the projected user need vector for user u. Dd is the projecteddesign vector for cell phone d. Tables 17 and 18 show Kansei vectorsand projected Kansei vectors for one user (User 39) and all 12Table 4 cell phones.

8.4. KE-LSA matching results

Table 19 shows user choice and KE-LSA choice for one user(User 39), user choices for all users, KE-LSA choices for all users,and KE-LSA accuracy. For one user, User 39 chose Phone 5 (bold).The KE-LSA approach chose Phone 12 (bold).


User 39 optimized CDA. User 39 chose the phone that matcheduser concept most accurately. The KE-LSA approach optimized co-sine to optimize UCDA. The KE-LSA approach tried to choose thephone that matched user needs most accurately. The KE-LSA ap-proach did optimize UCDA accurately. The KE-LSA approach didnot optimize CDA accurately, (p-values = 0.0000, 0.0000, 0.0000,0.0983).

For all users, users chose Phone 3 most often (bold). The KE-LSAapproach chose Phone 12 most often (bold). Users chose phoneswith higher CDA and lower UCDA (bold). Users chose phones thatmatched user concepts more accurately. The KE-LSA approachchose phones with higher UCDA and lower CDA (bold). The KE-LSA approach chose phones that matched user needs more accu-rately, (p-values = 0.0000) (bold).

The KE-LSA approach matched user choices, needs, and con-cepts with (12.7% CMA, 90.4% UCDA, 63.9% CDA). The KE-LSA ap-proach chose phones with higher UCDA than user choices. TheKE-LSA approach chose phones with lower CMA and CDA than userchoices, (p-values = 0.0000) (bold).

Table 26LSE projected user need and design descriptions.

User 1 2 3 4 5 6

0.28 �4.77 �4.45 �4.52 �4.77 �4.35 �4.61�1.76 3.62 �4.25 �3.53 2.71 �4.60 1.60

0.19 �2.87 �4.50 �3.72 �4.54 1.41 �0.44�0.11 �1.94 0.76 �6.09 �1.48 �3.33 0.48�0.03 �1.83 �4.03 3.63 0.06 4.34 �5.35

0.01 �0.23 0.45 4.61 �0.64 1.58 �7.420.00 1.71 0.26 1.82 2.50 3.56 �2.580.00 0.84 3.40 �1.10 �1.55 0.23 �1.040.00 4.36 3.18 �2.01 2.20 3.90 2.630.00 �3.32 �4.64 �2.38 3.52 �0.56 �0.950.00 �5.86 �0.34 5.39 �5.10 5.09 �3.060.00 �5.49 �7.68 3.35 �5.92 6.09 3.950.00 1.97 1.65 �2.75 4.23 �0.91 �4.680.00 �1.86 �0.55 1.98 �2.02 �1.55 3.340.00 5.26 2.20 1.31 4.70 �4.23 2.45�0.83 0.30 1.69 1.35 0.82 0.26 �0.06

0.05 0.75 0.02 �0.07 0.64 �0.40 0.58�0.93 0.79 0.65 1.02 0.90 0.50 0.45

Table 27LSE matching results.

User choice and LSE-M choice, one userPhone 11 12 10 5 7 2

Cos 0.268 0.216 0.205 0.182 0.174 0.13CMA 0.0 0.0 0.0 100.0 0.0 0.0UCDA 88.1 88.2 83.7 80.9 80.9 57.4CDA 62.6 44.0 80.9 99.4 52.9 43.3



LSE-M choice, all usersPhone 11 10 1 12 4 7


LSE-M accuracyApproach User KE-QT1 KE-LS

CMA 100.0 7.6 12.7UCDA 74.9 83.5 90.4CDA 98.4 64.1 63.9

The KE-LSA approach did improve UCDA, compared to the KE-QT1 approach. The KE-LSA approach did not improve CMA orCDA, compared to the KE-QT1 approach, (p-values = 0.0010,0.2946, 0.9226) (bold). The goal of this study is to improve match-ing accuracy, compared to the KE-QT1 approach.

KE-LSA accuracy depends upon KE-LSA semantic space dimen-sion. Table 20 shows case study KE-LSA CMA, UCDA, and CDA forsemantic space dimensions k = 2–6. The case study reduced KE-LSA semantic space dimension to k = 4. KE-LSA CMA was the sameat k = 4 and k = 5. KE-LSA UCDA was higher at k = 4.

9. The LSE approach

The LSE-M (matching) approach creates complete conceptualuser need descriptions, creates a complete conceptual LSE seman-tic space model, creates complete conceptual design descriptions,projects user need descriptions and design descriptions into theLSE semantic space, and matches projected user need descriptionsto projected design descriptions, in the semantic space.

7 8 9 10 11 12

�4.36 �4.52 �4.55 �4.33 �4.28 �4.28�4.15 0.67 2.97 �4.66 �5.58 �4.54

0.02 0.77 6.41 3.53 1.92 3.90�0.88 5.40 �1.20 �6.38 �2.03 �2.29�7.22 �1.31 0.82 3.59 2.26 �1.98�7.00 �1.84 1.03 3.62 �2.43 �1.02�3.58 6.71 �0.19 1.99 �5.05 �3.16

0.32 �3.36 �6.02 �0.43 3.12 �0.281.17 4.74 1.24 �3.13 0.43 3.14�3.77 �3.41 �3.76 �0.43 5.23 �1.37

3.71 �4.05 �1.35 4.50 3.38 �3.200.64 �0.79 �3.53 1.96 �5.11 �3.370.57 �4.88 �2.04 �4.02 �0.29 �2.372.35 �0.37 1.60 3.46 �3.48 2.730.91 1.09 6.12 0.31 �2.26 8.100.53 �0.14 �1.88 �0.28 0.21 �0.370.07 0.19 0.32 �0.39 �0.43 �0.250.37 0.11 0.12 0.49 0.42 0.23

3 8 6 9 4 1

0 0.094 �0.126 �0.150 �0.194 �0.259 �0.2970.0 0.0 0.0 0.0 0.0 0.0

54.9 7.5 16.5 36.6 13.8 11.980.2 51.9 42.4 42.7 32.7 32.6

6 1 2 8 7 9

3.8 2.5 2.5 2.5 1.3 0.0100.0 100.0 100.0 100.0 100.0

38.6 20.7 74.3 46.3 72.596.6 94.5 98.5 98.2 99.6

3 2 5 6 8 9

1.3 0.0 0.0 0.0 0.0 0.00.0 12.0 18.8

93.4 90.5 84.454.1 63.6 82.5

A LSE-M p-value p-value

15.2 0.0000 0.134986.9 0.0000 0.110368.4 0.0000 0.1083

Table 28LSE matching accuracy, k = 2–18.

Approach Dimension Information CMA UCDA CDA

KE-QT1 k = 6 100.0 7.6 83.5 64.1KE-LSA k = 4 92.7 12.7 90.4 63.9

LSE k = 2 15.4 7.6 84.7 61.5LSE k = 4 28.5 12.7 88.9 62.5LSE k = 6 41.4 12.7 89.1 62.7LSE k = 8 54.2 16.5 89.0 63.6LSE k = 10 67.1 13.9 89.0 62.9LSE k = 12 79.9 13.9 89.2 63.4LSE k = 14 92.8 13.9 89.1 62.9LSE k = 16 99.7 15.2 86.9 68.4LSE k = 18 100.0 15.2 85.9 68.0


The LSE-D (design) approach creates complete conceptualuser need descriptions, creates a complete conceptual LSEsemantic space model, projects user need descriptions into theLSE semantic space, and matches projected user need descrip-tions to LSE semantic space design elements, in the semanticspace (Table 2).


The LSE approach creates complete conceptual (aspirational,emotional, functional, physical) user need descriptions. The userneed descriptions can contain any combination of (aspirational,emotional, functional, physical) elements. In the case study, usersdid not provide (aspirational, functional, physical) user needs.Therefore, the case study user need descriptions only contain(emotional, physical) elements.

The case study LSE user need description, for any user, is a 27-element vector Du. The vector contains six emotional elements and21 physical elements. Each emotional element is derived from aSection 5.1 user need

Dui ¼ �1� ðuser needi � survey meanÞ ð11Þ

Table 29LSE design results.

User choice and LSE-D design, one userPhone New 5 10 11 3 12

Cos 0.429 0.429 0.429 0.429 0.238 0.238CMA 100.0 100.0 0.0 0.0 0.0 0.0UCDA 87.7 80.9 83.7 88.1 54.9 88.2CDA 99.4 99.4 80.9 62.6 80.2 44.0


Percent 24.1 22.8 16.5 11.4 8.9 3CMA 100.0 100.0 100.0 100.0 100.0 100UCDA 69.6 79.5 89.1 83.6 83.8 65CDA 98.1 98.9 98.9 98.6 98.8 98

LSE-D design, all usersPhone New New New New New New

Percent 24.1 22.8 16.5 11.4 8.9 3CMA 15.8 55.6 46.2 33.3 42.9 66UCDA 91.1 87.3 93.2 94.7 84.0 90CDA 98.3 76.3 87.4 86.5 85.5 78

LSE-DApproach User KE-QT1 KE-LSA

CMA 100.0 7.6 12.7UCDA 74.9 83.5 90.4CDA 98.4 64.1 63.9

Each physical element corresponds to a Table 5 design element.In the case study, users did not provide user needs for physical

design elements. Therefore, each physical element was set to ‘0’.Designers created LSE user need descriptions, for all 79 case studyusers. The case study LSE user need description for User 39 is[2,1,0,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0].

9.2. Create a model

The LSE approach uses SVD to create a complete conceptual LSEsemantic space model from complete conceptual design descrip-tions for sample designs (Table 21). The LSE approach uses theLSE model to project complete conceptual user need descriptionsand complete conceptual design descriptions into the LSE semanticspace.

In the case study, designers used SVD to create a conceptual(emotional, physical) LSE semantic space model from the Table 21conceptual design descriptions (‘conceptual design element (CDE)by conceptual design (CD) matrix’).

M0 ¼ E0S0D00 ð12Þ

E0 is a 27 � 18 matrix, S0 is an 18 � 18 diagonal matrix, and D0 is an18 � 18 matrix (Tables 22–24). Rows in E0 and D0 represent pro-jected design elements and projected design descriptions in the18-dimensional semantic space. The values in S0 represent scalingfactors, with respect to the 18 dimensions. The diagonal elementsin S0 are all positive and ordered in decreasing magnitude.

When used to project sample designs into and out of the LSEsemantic space, the LSE semantic space model matched the eigh-teen sample designs with 100.0% accuracy. The results show thatthe LSE semantic space model can create design descriptions with100.0% model accuracy (MA). The LSE approach can match designsto user needs with up to 100.0% matching accuracy (UCDA).


The LSE approach creates complete (aspirational, emotional,functional, physical) design descriptions. The LSE-M approach

9 8 2 4 7 1 6

0.238 0.048 0.048 0.048 �0.143 �0.143 �0.3330.0 0.0 0.0 0.0 0.0 0.0 0.0

36.6 7.5 57.4 13.8 80.9 11.9 16.542.7 51.9 43.3 32.7 52.9 32.6 42.4

6 1 2 8 7 9

.8 3.8 2.5 2.5 2.5 1.3 0.0

.0 100.0 100.0 100.0 100.0 100.0

.0 38.6 20.7 74.3 46.3 72.5

.5 96.6 94.5 98.5 98.2 99.6

New New New New New New

.8 3.8 2.5 2.5 2.5 1.3 0.0

.7 0.0 0.0 100.0 0.0 0.0

.3 90.0

.7 56.0

LSE-M LSE-D p-value p-value

15.2 36.7 0.0000 0.000086.9 89.6 0.0000 0.002968.4 73.0 0.0000 0.0050

Table 30Designs, one user.

Approach User KE-QT1 KE-LSA LSE-D

Phone 5 11 12 new design

CMA 100.0 0.0 0.0 100.0


chooses design elements, creates designs, creates design descrip-tions, projects user needs and designs into the LSE semantic space,matches projected needs to designs, and chooses the best designfor each user.

The LSE-D approach projects user needs into the LSE semanticspace, matches projected user needs to LSE semantic space designelements, and chooses the best design for each user. The LSE-D ap-proach creates new conceptual design descriptions from concep-tual user need descriptions.

In the case study, designers projected user needs and designsinto the LSE semantic space by multiplying each vector by transfor-mation matrices, according to Eq. (12). Designers matched pro-jected vectors, in the LSE semantic space, by calculating thecosine of the angle between the projected vectors

DuS � DdSjDuSj � jDdSj ð13Þ

Du and Dd are projected user needs and projected designs for a givenuser and a given design. Designers used the LSE-M approach tomatch designs to user needs and also used the LSE-D approach tomatch user needs to LSE semantic space design elements. Tables25 and 26 show user need, design, projected user need, and pro-jected design descriptions for one user (User 39) and all twelveTable 4 cell phone designs.

‘arc’

UCDA 80.9 88.1 88.2 87.7

CDA 99.4 62.6 44.0 99.4

‘elegant’ 3.24 2.95 3.22 3.02

‘simple’ 3.32 3.18 3.58 3.68

‘high tech’ 3.03 3.09 2.98 3.24

‘luxurious’ 4.08 3.86 3.54 3.40

‘beautiful’ 3.39 3.08 3.21 3.04

‘unique’ 4.05 3.97 3.60 4.18

‘type’

‘screen shape’

‘top shape’

‘body shape’

‘bottom shape’

‘number keys’

‘folding’

‘straight’

‘straight’

‘straight’

’25:16’

‘sliding’

‘sliding’

’25:16’

‘straight’

‘round’

‘sliding’

‘arc’

‘arc’

‘straight’

‘round’ ‘round’

’25:16’

‘straight’

‘straight’ ‘straight’

‘straight’ ‘arc’

‘4:3’

9.4. LSE-M matching results

Table 27 shows user choice and LSE-M choice for one user (User39), user choices for all users, LSE-M choices for all users, and LSE-M accuracy. For one user, User 39 chose Phone 5 (bold). The LSE-Mapproach chose Phone 11 (bold).

User 39 optimized CDA. User 39 chose the phone that matcheduser concept most accurately. The LSE-M approach optimized cosineto optimize CDA. The LSE-M approach tried to choose the phone thatmatched user concept most accurately. The LSE-M approach didoptimize UCDA accurately. The LSE-M approach did not optimizeCDA accurately, (p-values = 0.0000,0.0000,0.0000,0.0546) (bold).

For all users, users chose Phone 3 most often (bold). The LSE-Mapproach chose Phone 11 most often (bold). Users chose phoneswith higher CDA and lower UCDA. Users chose phones thatmatched user concepts more accurately. The LSE-M approachchose phones with higher UCDA and lower CDA. The LSE-M ap-proach chose phones that matched user needs more accurately,(p-values = 0.0000) (bold).

The LSE-M approach matched user choices, needs, and conceptswith (15.2% CMA, 86.9% UCDA, 68.4% CDA). The LSE-M approachchose phones with higher UCDA than user choices. The LSE-M ap-proach chose phones with lower CMA and CDA than user choices,(p-values = 0.0000) (bold).

The LSE-M approach did not improve UCDA, compared to theKE-QT1 approach. The LSE-M approach did not improve CMA orCDA, compared to the KE-QT1 approach, (p-val-ues = 0.1103,0.1349,0.1083) (bold). The goal of this study is to im-prove matching accuracy, compared to the KE-QT1 approach.

LSE-M accuracy depends upon LSE-M semantic space dimen-sion. Table 28 shows case study LSE-M CMA, UCDA, and CDA forsemantic space dimensions k = 2–18. The case study reduced LSE-M semantic space dimension to k = 16 (bold). LSE-M CMA washigher at k = 8. LSE-M CDA was higher at k = 16.

‘function keys’ ‘seven keys’ ‘five keys’ ‘five keys’ ‘seven keys’

9.5. LSE-D matching results

Table 29 shows user choice and LSE-D design for one user (User39), user choices for all users, LSE-D designs for all users, and LSE-D

accuracy. For one user, User 39 chose Phone 5 (bold). The LSE-Dcreated a new design that matched Phone 5 most accurately (bold).

User 39 optimized CDA. User 39 chose the phone that matcheduser concept most accurately. The LSE-D approach optimized co-sine to optimize CDA. The LSE-D approach tried to create a designthat matched user concept most accurately. The LSE-D approachdid optimize UCDA accurately. The LSE-D approach did optimizeCDA accurately, (p-values = 0.0000,0.0000,0.0183,0.0014).

For all users, users chose Phone 3 most often (bold). The LSE-Dapproach created a new design for each user (bold). Users chosephones with higher CDA and lower UCDA. Users chose phones thatmatched user concepts more accurately. The LSE-D approach cre-ated designs with higher UCDA and lower CDA. The LSE-D ap-proach created designs that matched user needs more accurately,(p-values = 0.0000) (bold).

The LSE-D approach matched user choices, needs, and conceptswith (36.7% CMA, 89.6% UCDA, 73.0% CDA). The LSE-D approach


created designs with higher UCDA than user choices. The LSE-D ap-proach created designs with lower CMA and CDA than user choices,(p-values = 0.0000) (bold).

The KE-QT1 approach matched user choices, needs, and con-cepts with (7.6% CMA, 83.5% UCDA, 64.1% CDA). The LSE-D ap-proach did improve UCDA, compared to the KE-QT1 approach.The LSE-D approach did improve CMA and CDA, compared to theKE-QT1 approach, (p-values = 0.0029, 0.0000, 0.0050) (bold). TheLSE-D approach achieves the goal of this study.

10. Case study designs

Table 30 shows user choice, KE-QT1 choice, KE-LSA choice, andthe LSE-D design for one user. The user (User 39) chose Phone 5.The KE-QT1 approach chose Phone 11. The KE-LSA approach chosePhone 12. The LSE-D approach created a new cell phone design thatmatched the user’s choice (Phone 5) most accurately.

Table 31Designs, all users.

Approach User User User LSE-D

Phone 3 5 10 new design

CMA 100.0 100.0 100.0 33.3

UCDA 69.6 79.5 89.1 88.7

CDA 98.1 98.9 98.9 68.7

‘elegant’ 3.68 3.24 3.16 3.28

‘simple’ 2.62 3.32 3.50 3.59

‘high tech’ 3.56 3.03 2.72 4.07 ‘luxurious’ 4.82 4.08 3.74 3.37

‘beautiful’ 3.99 3.39 3.36 4.57

‘unique’ 4.70 4.05 3.67 3.28

‘type’

‘screen shape’

‘top shape’

‘body shape’

‘bottom shape’

‘number keys’

‘function keys’

‘vertical’

‘straight’

‘straight’

‘sliding’

‘straight’

‘straight’

‘straight’

‘folding’

’25:16’

‘straight’

‘straight’

‘straight’ ‘straight’

‘straight’

‘vertical’

‘straight’

‘straight’

‘round’ ‘straight’ ‘straight’

‘seven keys’ ‘seven keys’ ‘seven keys’

‘4:3’ ‘4:3’ ‘1:1’

‘straight’

‘five keys’

The KE-QT1 choice matched user (needs, concept) with (88.1%UCDA, 62.6% CDA). The KE-LSA choice matched user (needs, con-cept) with (88.2% UCDA, 44.0% CDA). The LSE-D design matcheduser (needs, concept) with (87.7% UCDA, 99.4% CDA). The LSE-D de-sign improved CDA 36.8%, over the KE-QT1 choice.

The user choice matched user (needs, concept) with (80.9%UCDA, 99.4% CDA).The LSE-D design matched four design elementsin the user’s choice (Phone 5), ‘screen shape’, ‘top shape’, ‘bodyshape’, and ‘function keys’. The LSE-D approach chose ‘type’, ‘bot-tom shape’, and ‘number keys’ to improve UCDA. The LSE-D designimproved UCDA 6.8%, compared to the user’s choice.

Table 31 shows the top three user choices and the top LSE-D de-sign, for all users. Most users chose Phones 3, 5, and 10. The LSE-Dapproach created new designs for each user. The top user choicematched user (needs, concepts) with (69.6% UCDA, 98.1% CDA).The top LSE-D design matched user (needs, concepts) with (88.7%UCDA, 68.7% CDA).

Most users chose similar phones. The top three user choicesonly varied by ‘type’ and ‘screen shape’. The LSE-D design matchedfour design elements in the top user choice (Phone 3), ‘type’, ‘topshape’, ‘body shape’, and ‘number keys’. The LSE-D approach chose‘screen shape’, ‘bottom shape’, and ‘function keys’ to improveUCDA. The LSE-D design improved UCDA 19.1%, compared to thetop user choice.

Case study results show that the LSE-D approach models con-ceptual design processes more accurately than other approaches.Case study results show that the LSE-D improves model accuracy(MA) and matching accuracy (CMA, UCDA, and CDA), comparedto other approaches. The LSE-D creates design descriptions moreaccurately than other approaches. The LSE-D approach matches de-signs to user choices, needs, and concepts more accurately thanother approaches.

11. Conclusions

User-centered design (UCD) plays a vital role in the productdevelopment process. UCD approaches match designs to userneeds. Matching designs to needs improves overall product quality,customer satisfaction, and product success. However, the matchingprocess is inherently difficult and inaccurate. Matching accuracy isgenerally very low. As a result, most designs still do not match userneeds.

The goal of this study is to improve matching accuracy. Toachieve the goal, this study introduces conceptual UCD and LatentSemantic Engineering (LSE), a new conceptual UCD approach, de-fines measures for model accuracy (MA), conceptual matchingaccuracy (CMA), user-centered design accuracy (UCDA), and con-ceptual design accuracy (CDA), and compares the LSE approachto other approaches.

The LSE approach models conceptual design processes moreaccurately than other approaches. The LSE approach creates com-plete conceptual (aspirational, emotional, functional, and physical)designs from complete conceptual user needs. The LSE approachuses a LSE semantic space model to create designs from user needs.The LSE approach improves MA and CMA, UCDA, and CDA, com-pared to other approaches.

This study uses a case study to compare the LSE approach to KE-LSA and KE-QT1 approaches. The case study uses the KE-QT1, KE-LSA, and LSE approaches to create customized cell phone designsfor individual users. In the case study, all three approaches usethe same set of user needs to create user need descriptions. Allthree approaches use the same set of sample design rankings tocreate models.

All three approaches use different techniques to create userneed descriptions, models, and design descriptions. The case study


determines MA, CMA, UCDA, and CDA for all three approaches. Thecase study uses matching to determine MA and CMA. The casestudy uses vector cosine matching to determine UCDA and CDA.The case study converts matching results to accuracy values, inpercent.

In the case study, designers identified user needs and userchoices by completing an Internet survey. Users, Kansei words,and phones had significant impacts on user needs and user choices,(p-values = 0.0000). The survey controlled user response (withKansei words), user interaction with the product (with images),and age range. Results can be generalized to users, in the targetage range, that respond to Internet surveys.

In the case study, designers created sample design rankings byconducting an Internet survey. Sample designs and Kansei wordshad significant impacts on rankings (p-values = 0.0002,0.0048).The survey controlled age range and demographic ratios to removedemographic bias from the rankings. Results can be used to createmodels for users, in the target age range, that respond to Internetsurveys.

The KE-QT1, KE-LSA, and LSE approaches created designdescriptions with (87.4%, 87.4%, 100.0%) MA. The KE-QT1 and KE-LSA approaches did not create design descriptions accurately. TheLSE model created design descriptions accurately. The LSE modelimproved MA 12.6%, compared to the KE-QT1 and KE-LSA models.

The KE-QT1, KE-LSA, and LSE-D (design) approaches matcheddesigns to user choices with (7.6%, 12.7%, and 36.7%) CMA. TheKE-QT1 and KE-LSA approaches did not match designs to userchoices accurately. The LSE-D approach matched designs to userchoices more accurately than other approaches. The LSE-D ap-proach improved CMA 29.1%, compared to the KE-QT1 approach.

The KE-QT1, KE-LSA, and LSE-D (design) approaches matcheddesigns to user needs with (83.5%, 90.4%, and 89.6%) UCDA. TheKE-QT1 approach did not match designs to user needs accurately.The LSE-D (design) approach matched designs to user needs moreaccurately than the KE-QT1 approach. The LSE-D approach im-proved UCDA 6.1%, compared to the KE-QT1 approach.

The KE-QT1, KE-LSA, and LSE-D (design) approaches matcheddesigns to user concepts with (64.1%, 63.9%, and 73.0%) CDA. TheKE-QT1 and KE-LSA approaches did not match designs to user con-cepts accurately. The LSE-D (design) approach matched designs touser concepts more accurately than other approaches. The LSE-Dapproach improved CDA 8.9%, compared to the KE-QT1 approach.

The LSE approach models conceptual design processes moreaccurately than other approaches. The LSE approach improvesMA and CMA, UCDA, CDA, compared to other approaches. TheLSE approach creates design descriptions more accurately thanother approaches. The LSE approach matches designs to userchoices, needs, and concepts more accurately than otherapproaches.

The LSE-D approach is an optimal approach. The LSE-D ap-proach creates complete conceptual user need descriptions, createsa minimum-size semantic space model, projects user need descrip-tions into the semantic space with 100% accuracy, matches userneeds to design elements by minimizing vector cosine, creates anoptimal design, and achieves maximum UCDA at maximum CDA.

This study uses the LSE approach to create conceptual (emo-tional, physical) designs from conceptual (emotional, physical)user needs. The LSE approach can also be used to create conceptual

designs that contain any combination of conceptual (aspirational,emotional, functional, physical) design elements, from any combi-nation of conceptual (aspirational, emotional, functional, physical)user needs.

The case study uses the LSE approach to create individually cus-tomized traditional cell phone designs. The LSE approach can alsobe used to create other conceptual designs for either individual ormultiple users. For example, the LSE approach can be used to cre-ate screen functions for smart phones, handheld computers, orPDAs, from any combination of conceptual user needs.

This study introduces conceptual UCD and LSE, a conceptualUCD approach, defines MA, CMA, UCDA, and CDA, and determinesMA, CMA, UCDA, and CDA for KE-QT1, KE-LSA, LSE-M, and LSE-Dapproaches. The study improves MA and CMA, UCDA, and CDA.Study results can be used to improve product quality, customersatisfaction, product success, and the product developmentprocess.

Acknowledgement

National Science Council, Taiwan (NSC 100-2221-E-002-058)provided support.

References

[1] C.C. Chen, M.C. Chuang, Integrating the Kano model into a robust designapproach to enhance customer satisfaction with product design, Int. J. Prod.Econ. 114 (2008) 667–681.

[2] M. Nagamachi, Kansei engineering: a new ergonomic consumer-orientedtechnology for product development, Int. J. Ind. Ergon. 15 (1995) 3–11.

[3] G. Kannan, Implementation of fuzzy quality function deployment in anautomobile component to improve the quality characteristics, Qual. Eng. 20(2008) 321–333.

[4] S. Barone, A. Lombardo, P. Tarantino, A weighted logistic regression forconjoint analysis and Kansei engineering, Qual. Reliab. Eng. Int. 23 (2007) 689–706.

[5] M. Ohkura, S. Goto, T. Aoto, Systematic study for ‘kawaii’ products (fifthreport): relation between kawaii feelings and biological signals, ICCAS-SICE –ICROS-SICE Int. Joint Conf. Fukuoka Japan, 2009.

[6] H. Lai, Y. Lin, C. Yeh, C. Wei, User-oriented design for the optimal combinationon product design, Int. J. Prod. Econ. 100 (2006) 253–267.

[7] K. Fukushima, H. Kawata, Y. Fujiwara, G. Hirokazu, Human sensory perceptionoriented image processing in a color copy system, Int. J. Ind. Ergon. 15 (1995)63–74.

[8] R. Koch, The 80/20 Principle, Doubleday, New York, 1998.[9] P.W. Jordan, How to Make Brilliant Stuff that People Love and Make Big Money

Out of It, John Wiley and Sons, Chichester, 2002.[10] R. Seva, M. Helander, The influence of cellular phone attributes on users’

affective experiences: a cultural comparison, Int. J. Ind. Ergon. 39 (2009) 341–346.

[11] H. Khalid, M. Helander, Customer emotional needs in product design,Concurrent Eng. 14 (3) (2006) 197–206.

[12] S. Deerwester, S. Dumais, G. Furnas, T. Landauer, R. Harshman, Indexing bylatent semantic analysis, J. Am. Soc. Inform. Sci. 41 (1990) 391–407.

[13] B. Rosario, Latent semantic indexing: an overview, Infosys 240 (2000).[14] M. Berry, S. Dumais, G. O’Brien, Using linear algebra for intelligent information

retrieval, SIAM Rev 37 (4) (1995) 573–595.[15] D. Norman, Design of Everyday Things, MIT Press, London, 1988.[16] C.H. Hsu, T.M. Chang, S.Y. Wang, P.Y. Lin, Integrating Kano’s model into Quality

Function Deployment to facilitate decision analysis for service quality, in: Proc.of the 8th Int. Conf. on Math. and Comp. Business and Econ., Vancouver,Canada, 2007, pp. 226–232.

[17] C. Ling, W. Hwang, G. Salvendy, A survey of what customers want in a cellphone design, Behav. Inform. Tech. 26 (2) (2007) 149–163.

[18] K. Chen, S.H. Chiu, F.C. Lin, Kansei Design with Cross Cultural Perspectives,Usability and Internationalization, Springer-Verlag, Berlin, 2007. pp. 47–56.

[19] M. Nagamachi, Kansei engineering as a powerful consumer-orientedtechnology for product development, Appl. Ergon. 33 (2002) 289–294.