Prospect-Theoretic Modeling of Customer Affective-Cognitive Decisions.pdf

16
468 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 44, NO. 4, AUGUST 2014 Prospect-Theoretic Modeling of Customer Affective-Cognitive Decisions Under Uncertainty for User Experience Design Feng Zhou, Yangjian Ji, and Roger J. Jiao Abstract—In order to incorporate both affective and cognitive factors in the decision-making process, a user experience (UX) eval- uation function based on cumulative prospect theory is proposed for three different affective states and two different types of prod- ucts (affect-rich versus affect-poor). In order to tackle multiple parameters involved in the UX evaluation function, a hierarchi- cal Bayesian model is proposed with a technique called “Markov chain Monte Carlo.” It estimates parameters that represent differ- ent cognitive tendencies and affective influences for customers at the individual and group levels by generating posterior probability density functions of the parameters to incorporate inherent uncer- tainty. An experiment with four hypotheses was designed to test the proposed model. We found that: 1) anxious participants tend to be more risk-averse than those in joy and excitement; 2) joyful and excited participants tend to be more risk-seeking than those in anxiety in UX-related choice decision making; 3) all participants tend to be averse to unpleasant UX; and 4) participants tend to value by feeling for affect-rich products and value by calculation for affect-poor products. Furthermore, the models of five different types can predict choice decision making between product pro- files with around 80% accuracy. In summary, the results explain affective-cognitive decision-making behavior in the complex do- main of UX design and, thus, illustrate the potential and feasibility of the proposed method. Index Terms—Affective-cognitive decision making, cumulative prospect theory (CPT), customer preference, hierarchical Bayesian model, user experience (UX) design. I. INTRODUCTION A MODERN product, such as iPhone or iPad, works not only because of its self-contained functionality, but also owing to its hedonic user experience (UX)—it delights customers [1]! Looking at the whole spectrum of UX, an understanding of the flow patterns and directions of customer preference, per- ception, and experience is essential to getting “the whole thing” Manuscript received June 13, 2013; revised September 3, 2013, October 31, 2013 and March 4, 2014; accepted April 13, 2014. Date of publication May 12, 2014; date of current version July 11, 2014. This paper was recommended by Associate Editor C. Wu. This work was surported in part by the National Natural Science Foundation of China under Project 51275456. F. Zhou and R. J. Jiao are with the George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail: [email protected]; [email protected]). Y. Ji is with the Industrial Engineering Centre, Department of Mechan- ical Engineering, Zhejiang University, Hangzhou 310027, China (e-mail: [email protected], Corresponding author). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/THMS.2014.2318704 right. Human perception on UX originates from the evolution of customers’ affective states triggered by stimuli (events) along a chain of executing cognitive tasks involved in the human– product interaction process [2]. Therefore, UX has two essential aspects: affect and cognition. Engineering design traditionally emphasizes products’ functional requirements, yet with limited consideration of customers’ affective and cognitive needs [3]. Human affect plays a significant and useful role in human deci- sion making [4]. It is, therefore, imperative for design research to bring in “the human dimension.” In this paper, mental processes, such as sensation, memory, attention, perception, and problem solving, refer to aspects of cognition. They described as cognitive or the cognitive system (i.e., the “analytic system”) are thought to be necessary to per- form a cognitive task, such as decision making [5]. The analytic system makes use of conscious deliberate cognitive processes with various algorithms and normative rules that produce logical behavior and maximize expected utility [6]. Thus, the operations of the cognitive system are typically slower, more effortful, and more likely deliberately controlled, e.g., the process of solv- ing a mathematical question. On the other hand, affect is an encompassing term, including emotions, feelings, moods, and evaluations. An affective state is often a transient emotion, such as fear of a situation, which influences decision making [7]. The affective system, which is also known as the “experien- tial system,” employs past experiences, emotion-related associ- ations, and intuitions for decision making [6]. The operations of the affective system are often fast, automatic, effortless, and associative. Users tend to be more susceptible to affect-rich (i.e., hedonic) products than affect-poor (i.e., utilitarian) prod- ucts. Affect-rich products are those that allow the consumer to feel pleasure, fun, and enjoyment from buying and using them, whereas affect-poor products are purchased for their practical and functional uses [8]. For example, shoppers often experi- ence impulsive buying for affect-rich products (e.g., a favorite music album) rather than for affect-poor products (e.g., a com- puter software CD). In addition, the impulsive buying process is considered extraordinary, exciting, and spontaneous [9]. We argue that the affective system and the cognitive need to work collaboratively in order to make the best decisions [10]. While affective elements and subjective experience are well known to influence human decision making, prevailing compu- tational models for analyzing and simulating human perception and evaluation on UX are mainly cognition-based models [12]. Expected utility theory assumes that humans make decisions based on a deliberate cost-benefit analysis [13]. Recent models 2168-2291 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Transcript of Prospect-Theoretic Modeling of Customer Affective-Cognitive Decisions.pdf

468 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 44, NO. 4, AUGUST 2014

Prospect-Theoretic Modeling of CustomerAffective-Cognitive Decisions Under Uncertainty

for User Experience DesignFeng Zhou, Yangjian Ji, and Roger J. Jiao

Abstract—In order to incorporate both affective and cognitivefactors in the decision-making process, a user experience (UX) eval-uation function based on cumulative prospect theory is proposedfor three different affective states and two different types of prod-ucts (affect-rich versus affect-poor). In order to tackle multipleparameters involved in the UX evaluation function, a hierarchi-cal Bayesian model is proposed with a technique called “Markovchain Monte Carlo.” It estimates parameters that represent differ-ent cognitive tendencies and affective influences for customers atthe individual and group levels by generating posterior probabilitydensity functions of the parameters to incorporate inherent uncer-tainty. An experiment with four hypotheses was designed to testthe proposed model. We found that: 1) anxious participants tendto be more risk-averse than those in joy and excitement; 2) joyfuland excited participants tend to be more risk-seeking than those inanxiety in UX-related choice decision making; 3) all participantstend to be averse to unpleasant UX; and 4) participants tend tovalue by feeling for affect-rich products and value by calculationfor affect-poor products. Furthermore, the models of five differenttypes can predict choice decision making between product pro-files with around 80% accuracy. In summary, the results explainaffective-cognitive decision-making behavior in the complex do-main of UX design and, thus, illustrate the potential and feasibilityof the proposed method.

Index Terms—Affective-cognitive decision making, cumulativeprospect theory (CPT), customer preference, hierarchical Bayesianmodel, user experience (UX) design.

I. INTRODUCTION

AMODERN product, such as iPhone or iPad, works not onlybecause of its self-contained functionality, but also owing

to its hedonic user experience (UX)—it delights customers [1]!Looking at the whole spectrum of UX, an understanding ofthe flow patterns and directions of customer preference, per-ception, and experience is essential to getting “the whole thing”

Manuscript received June 13, 2013; revised September 3, 2013, October 31,2013 and March 4, 2014; accepted April 13, 2014. Date of publication May12, 2014; date of current version July 11, 2014. This paper was recommendedby Associate Editor C. Wu. This work was surported in part by the NationalNatural Science Foundation of China under Project 51275456.

F. Zhou and R. J. Jiao are with the George W. Woodruff School of MechanicalEngineering, Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail:[email protected]; [email protected]).

Y. Ji is with the Industrial Engineering Centre, Department of Mechan-ical Engineering, Zhejiang University, Hangzhou 310027, China (e-mail:[email protected], Corresponding author).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/THMS.2014.2318704

right. Human perception on UX originates from the evolution ofcustomers’ affective states triggered by stimuli (events) alonga chain of executing cognitive tasks involved in the human–product interaction process [2]. Therefore, UX has two essentialaspects: affect and cognition. Engineering design traditionallyemphasizes products’ functional requirements, yet with limitedconsideration of customers’ affective and cognitive needs [3].Human affect plays a significant and useful role in human deci-sion making [4]. It is, therefore, imperative for design researchto bring in “the human dimension.”

In this paper, mental processes, such as sensation, memory,attention, perception, and problem solving, refer to aspects ofcognition. They described as cognitive or the cognitive system(i.e., the “analytic system”) are thought to be necessary to per-form a cognitive task, such as decision making [5]. The analyticsystem makes use of conscious deliberate cognitive processeswith various algorithms and normative rules that produce logicalbehavior and maximize expected utility [6]. Thus, the operationsof the cognitive system are typically slower, more effortful, andmore likely deliberately controlled, e.g., the process of solv-ing a mathematical question. On the other hand, affect is anencompassing term, including emotions, feelings, moods, andevaluations. An affective state is often a transient emotion, suchas fear of a situation, which influences decision making [7].The affective system, which is also known as the “experien-tial system,” employs past experiences, emotion-related associ-ations, and intuitions for decision making [6]. The operationsof the affective system are often fast, automatic, effortless, andassociative. Users tend to be more susceptible to affect-rich(i.e., hedonic) products than affect-poor (i.e., utilitarian) prod-ucts. Affect-rich products are those that allow the consumer tofeel pleasure, fun, and enjoyment from buying and using them,whereas affect-poor products are purchased for their practicaland functional uses [8]. For example, shoppers often experi-ence impulsive buying for affect-rich products (e.g., a favoritemusic album) rather than for affect-poor products (e.g., a com-puter software CD). In addition, the impulsive buying processis considered extraordinary, exciting, and spontaneous [9]. Weargue that the affective system and the cognitive need to workcollaboratively in order to make the best decisions [10].

While affective elements and subjective experience are wellknown to influence human decision making, prevailing compu-tational models for analyzing and simulating human perceptionand evaluation on UX are mainly cognition-based models [12].Expected utility theory assumes that humans make decisionsbased on a deliberate cost-benefit analysis [13]. Recent models

2168-2291 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

ZHOU et al.: PROSPECT-THEORETIC MODELING OF CUSTOMER AFFECTIVE-COGNITIVE DECISIONS 469

based on behavioral decision theories focus on cognitive er-rors and heuristics in human judgments, but still ignore therole of emotion in human decision making [14]. Such a singlecognitive perspective is not optimal for analyzing human deci-sions toward UX, in which users’ affective states experiencedat the time of decision making often influence their perceptionsand choices [15]. The intimate coupling of affective and cogni-tive decisions has driven recent consensus on the integration ofemotion and cognition [16]. Several computational mechanismshave emerged, which treat cognition as a necessary antecedentto emotion [17]. However, the computational realizations ofaffect-cognition integration have largely been pragmatic, andthe link between core cognitive functions and emotions has yetto be fully explored [18].

In this research, the contributions can be summarized in threeaspects: We 1) propose a UX decision-making model based oncumulative prospect theory (CPT) that is more accurate thanexpected utility theory; 2) incorporate the influence of affect inthe UX model as an extension to the original CPT theory; and3) estimate parameters involved in the model as a way to testthe hypotheses and accommodate both individual heterogeneityand group homogeneity.

First, we adopt CPT as the basic decision-making UX model.Prospect theory [19] was originally proposed to describe de-cision making in the domain of behavioral economics. Unlikenormative decision models (e.g., expected utility theory), CPTis a descriptive decision model that describes how humans actu-ally make decisions rather than they “should.” A CPT valuefunction is defined with respect to a reference point, ratherthan in terms of an absolute value as in expected utility the-ory, and thus is reference dependent (see (1)). Such an emphasison the reference point conforms to the human perceptual pro-cess, which tends to notice shifts more than resting on staticstates [20]. Furthermore, CPT modifies the original prospecttheory by applying the probability distortions to the cumulativeprobabilities so that stochastic dominance is not violated. TheUX design essentially shares the same choice decision-makingprocess with those in behavioral economics. Customers choosedifferent product profiles, which result in different UX (unpleas-ant, neutral, or unpleasant). Thus, CPT provides a “legitimate”method to evaluate UX of a particular product profile. Preferencechoice decision making involves uncertainty, where the choiceprobability distributions of various design attributes favored bydifferent customers are unknown, although historical marketingdata imply certain information about such uncertainty. Amongthem, the most frequently chosen product attributes contributeto pleasant UX, and vice versa.

Second, we incorporate the influence of affect in the decision-making process by shaping the parameters involved in the CPT-based UX model. The original CPT model does not incorporatethe influence of affect. However, according to Ahn [12], the pa-rameters of the value function change systematically in sequen-tial decision-making situations, involving incidental affectivestates and task-related confidence. Therefore, it is possible to in-corporate affective factors by shaping the CPT parameters in theUX design, whereby human choice behavior is made regardingmultiple design attributes; the attribute corresponding to neutral

UX can be regarded as a reference point; positive and negativeUX functions can be formulated to express diverse customerpreferences; and affective influences and cognitive tendenciescan be incorporated in the value function and weighting functionby parameter shaping.

Third, in order to estimate multiple parameters associatedwith the UX model, a hierarchical Bayesian parameter esti-mation procedure is developed, taking affective influence intoaccount. The prevailing method to estimate model parameters iseither to average data across groups of participants to uncoverunderlying patterns or to use individual data to accommodateindividual differences. However, the former cannot accommo-date individual differences and the latter usually has few andthus noisier data, based on which the result is more unreliablecompared with the former (see, e.g., [21]). In this paper, a hier-archical Bayesian model is proposed to estimate the parametersinvolved in the UX model. It offers a principled and compre-hensive way to relate psychological models to experimental andobservational data [22]. It can identify how the variables arerelated, inferring causal influences between UX and product at-tributes that go beyond regression or correlation analysis [23].Meanwhile, a hierarchy of submodels is useful to analyze thedata with one model at the group level for studying customer ho-mogeneity and another model at the individual level for studyingcustomer heterogeneity.

The remainder of this paper proceeds as follows. InSection II, related work leading to the current research isreviewed. Section III presents the proposed UX model based onCPT. It gives an overview of the model architecture with fourphases, including a perceptual phase, an affective-cognitive rea-soning phase, a learning phase, and an evaluation phase. Theseare described in detail in Sections III-B to III-E. Section IVreports the empirical study for model parameter estimation andhypothesis testing. Estimation results and validation are pre-sented in Section V. We discuss the results and conclude thispaper in Sections VI and VII, respectively.

II. RELATED WORK

A. User Experience Modeling

Generally, there are two kinds of UX models existing in theliterature: qualitative and quantitative. Qualitative UX modelsare prevalent in the areas of human factors, human–computerinteraction, information engineering, and consumer psychology.These models usually identify the measurements and structuresof UX [24]. In measurement models, UX constructs are re-garded as latent variables and each has one or more manifestvariables. Data about the manifest variables are collected totest the measurement model using statistical techniques, such asanalysis of variance (ANOVA), factor analysis, and correlationand regression analysis. For example, Hassenzahl [25] exploredthe interplay of beauty, goodness, and usability in interactiveproducts. van Schaik and Ling [26] modeled UX for websiteswhere relationships among usability, hedonic value, beauty,and goodness were investigated. Another trend of measure-ment models is to relate physiological and behavioral measuresto UX in real time. These measures include facial expression,

470 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 44, NO. 4, AUGUST 2014

EMG, EEG, skin conductance response, and so on (see, e.g., [27]and [28]). A well-tested measurement model enables UX mea-sures to be meaningful and validated, and the nature of UXconstructs can inform how UX is operationalized and mea-sured [24]. Subsequently, a structural model can be used tofind out the cause-and-effect relations between UX constructsusing statistic techniques, such as ordinary least squares (see,e.g., [26]), structural equation modeling (see, e.g., [29]), andCohen’s path analysis method [30]. Partala and Kallinen [31]identified the most satisfying and unsatisfying UXs by studyingemotions, psychological needs, and contexts. Zhou et al. [32]studied the causal relations between UX constructs and productfeatures with fuzzy reasoning Petri nets.

UX needs to be quantified in order to develop UX measure-ment models. If this is done accurately, decision makers can eas-ily estimate the potential value of a product or service [24], [33].However, only a few quantitative models have been proposed.The simplest quantitative models were linear, with a weightedsum of different constructs of UX identified (e.g., usability, af-fect, user value, quality, and user satisfaction), and coefficientsestimated using statistical analysis, such as multiple regressionanalysis and maximum likelihood estimation. For example, Kimand Han [34] developed a quantitative usability index of con-sumer electronic products with a weighted sum of multiple di-mensions of usability. However, UX, especially emotional re-sponses to products, seldom exhibits linear characteristics, butrather is fuzzy and vague in nature [35]. Hence, complex mod-els, such as quadratic and S-shaped models, were developed.For example, Park et al. [33] compared five different models ofUX with three constructs (i.e., usability, affect, and user value)and found that quadratic and S-shaped UX models performedbest in a mobile device case study. The S-shaped UX modelwas based on prospect theory. However, they kept the originalparameters of prospect theory from behavioral economics [19],in which α = β = 0.88, and λ = 2.25, without consideringits applicability to the domain of UX. Therefore, it is impera-tive to have an appropriate method to estimate parameters ofquantitative UX models.

B. Affective Influence on Decision Making

Recent affective neuroscience and psychological studies havereported that human affect and emotional experience play a sig-nificant role in human learning and decision making [12]. Ahnand Picard [15] proposed an affective-cognitive decision frame-work for learning and decision making. Zhang and Liu [36] pre-sented a navigation system based on affective-cognitive learn-ing and decision making, which sped up the learning processand improved the capability of autonomous navigation. Poweret al. [37] emphasized the role of emotion in combination withhealth behavior models to provide a framework for conceptu-alizing patient decisions. Bracha and Brown [11] proposed anaffective decision-making model of choice under risk and un-certainty, and they posited that observed choices were the resultof a rational and emotional interaction process with a Nashequilibrium. Penolazzi et al. [38] studied the role of impul-sivity and reward sensitivity in affective and deliberative risky

decision making, and suggested that personality traits differen-tially alter decision-making behavior due to interactions with thedecision-making context. Bechara et al. [39] showed that covertbiases related to previous emotional experience of comparablesituations assisted the reasoning process, and facilitated the effi-cient processing of knowledge and logic necessary for consciousdecisions.

C. Decision-Based Design

Unlike traditional design that focuses on the designer’s per-spective, customer preferences in terms of expected utilitiesare the primary focus in the decision-based design [40]. Thedecision-based design recognizes the substantial role of deci-sions in design characterized by ambiguity, uncertainty, risk, andtradeoff [41]. Utility analysis is often used to build mathematicalmodels of a decision maker’s preference as a way to identify theoptimal option [42]. For example, Orsborn et al. [43] modeledpreferences for aesthetic forms using a utility function quanti-tatively. However, not much research work has involved prefer-ence modeling in terms of UX. For example, Zhou et al. [35]modeled customer preferences in terms of affective needs inaccordance with customer emotional satisfaction. Furthermore,normative decision models (e.g., expected utility theory) havefailed to explain behavior and decision making in real-world sit-uations. Descriptive decision models (e.g., CPT) describe howhumans make decisions as they are rather than as they shouldand thus are more appropriate to model human behavior for UXdesign and other real-life situations. For example, Xu et al. [44]applied CPT to model a traveler’s route choice decision makingand found that the results were more coherent with the experi-ment data than those obtained from the route choice model basedon expected utility theory. Likewise, Chow et al. [45] appliedprospect theory to estimate a discrete choice model for choos-ing a high-occupancy-vehicle lane or not based on the currentand perceived speeds. Compared with the binary logit model,they discovered that prospect-theoretic-based models were moreconsistent with empirical data.

III. PROPOSED USER EXPERIENCE MODEL

A. Overview of the Model Architecture

In general, UX is evolved as users’ affective-cognitive deci-sion making through their interactions with a variety of designattributes, denoted as a set, A = {ai}I , where I is the totalnumber of design attributes. These design attributes embodythe key characteristics of a product or service system. Each de-sign attribute may assume a number of levels, either discrete orcontinuous, A∗

i = {a∗ik}I×Li

, 1 ≤ k ≤ Li , where Li is the totalnumber of levels (instances) of ai , and k denotes the kth levelof ai . For example, the interior color of an aircraft cabin canbe a design attribute and may assume five attribute levels (e.g.,blue, orange, green, pink, and white). The perceived UX of aproduct or service is a holistic impression resulting from com-plex cognitive and affective interactions with the product profileformed by different design attribute levels [46], comprising afinite set, X = {xik}I×Li

, in the user’s mental space. While

ZHOU et al.: PROSPECT-THEORETIC MODELING OF CUSTOMER AFFECTIVE-COGNITIVE DECISIONS 471

Fig. 1. Architecture of the UX model based on CPT.

xik indicates a quantitative measure of UX for a specific designattribute level, X is the holistic measure of UX for the entiredesign. With regard to various attribute levels, it is importantthat users are able to make decisions based on their perceivedUX. Therefore, the problem of UX design in the context ofaffective-cognitive decision making is formulated:

Given: Design attributes A = {ai}I and attribute levels A∗i =

{a∗ik}I×Li

, 1 ≤ k ≤ Li ;Find: Optimal configuration of the product;Maximize: Holistic perceived UX for a product configuration,

i.e., U = X (a,p);Subject to:1) Affective influence on choice decision making,2) Cognitive influence on choice decision making, and3) Uncertainty involved in choice decision making,

where X(a,p) is an aggregation function that computes the holis-tic UX for a product profile a formed by specific design attributelevels with their choice probability vector p.

In order to answer the problem formulated above, we proposea UX model based on CPT. The architecture is illustrated inFig. 1. The model utilizes experiment data to project the shapeof the CPT value function and the weighting function in orderto deal with future decision making. The model comprises fourconsecutive phases, namely the perceptual phase, the affective-cognitive reasoning phase, the learning phase, and the evaluationphase. It assumes that the decision-making process is influencedby the user’s affective states, cognitive tendencies, risk attitudes,and product types when the decision making is about to happen.In the perceptual phase, the user identifies the product typesand estimates the reference point corresponding to neutral UX.In the affective-cognitive reasoning phase, the CPT-based valuefunction and the weighting function are formulated to evaluateUX for diverse design attribute levels. In the learning phase,shape parameters involved in the UX model are estimated basedon a hierarchical Bayesian model. In the evaluation phase, aCPT-based UX function is used for UX evaluation. These fourphases are described in detail below.

Fig. 2. CPT-based UX value functions.

B. Perceptual Phase

We assume that the design attributes and their levels are given.Then, the user’s perceived UX of a design attribute level a∗

ik canbe defined by a subjective value function v(a∗

ik ) weighted byhis or her subjective probability. In the perceptual phase, theperceived UX of various options is identified relative to a cer-tain design attribute level a∗

i,ref that gives a neutral UX andacts as a reference point. Hassenzahl and Tracinsky [47] pointout that UX involves dynamic context-dependent internal statesof users, including both instrumental and emotional aspects. Itis, hence, likely that the reference point varies among differentrespondents. To hedge against this problem, we set up individ-ual reference points for individual UX models to accommodatecustomer heterogeneity and take a grand mean as the referencepoint for all the customers within one market segment to ac-commodate customer homogeneity.

C. Affective-Cognitive Reasoning Phase

1) Subjective Value Function: CPT addresses important sub-jective influences on human choice decision making in the UXdesign using a value function v as follows [19]:

vik = v (a∗ik ) =

{(Δa∗

ik )α , Δa∗ik ≥ 0

−λ (−Δa∗ik )β , Δa∗

ik < 0(1)

where Δa∗ik = a∗

ik − a∗i,ref is the difference between the target

design attribute level a∗ik (eliciting (un)pleasant UX) and the

reference attribute level a∗i,ref (eliciting neutral UX). The CPT

subjective value function is defined with respect to a referencepoint and thus is reference dependent. In addition, α and β areparameters between 0 and 1, modulating the curvature of thesubjective value function, which represents a decision maker’ssensitivity to, risk attitude to, and affective influence on UX.Four aspects below indicate how the value function accountsfor the influence of affect and cognition on UX (see Fig. 2).

First, users’ cognitive appraisal of product attributes playsa significant role in assessing perceived UX. According to theappraisal theory, Ellsworth and Scherer [48] stated that humanusers evaluate product attributes (i.e., stimulus) in terms of per-ceived significance relative to needs and goals of the personconcerned. Therefore, the more conducive the product is toachieving users’ goals and satisfying his/her needs, the morepleasant the perceived UX would be. When predicted emotions

472 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 44, NO. 4, AUGUST 2014

(cognitive beliefs) of future outcomes and/or genuine emotionselicited by the product attributes are very positive, impulsivebuying may occur [9].

Second, the subjective value function has a diminishing sensi-tivity (i.e., 0 < α, β < 1), i.e., the change in UX decreases as thedistance between the reference point and the target design at-tribute level increases (see Fig. 2). Two psychological processesin constructing preferences are distinguished, i.e., valuation byfeeling and valuation by calculation [49]. Valuation by feelingfor affect-rich stimulus is scope-insensitive. For example, inDesvousges et al.’s study [50], the mean values of donation tosave 2000, 20 000, or 200 000 migrating birds from drowningin oil ponds were $80, $78, and $88, respectively. This wasexplained by Kahneman et al. [51] that Desvousges et al.’squestions evoked an affect-rich mental representation of an ex-hausted bird, its feathers soaked in black oil, unable to escape.The money the participants decided to donate was based on theiraffective reactions to this image. Valuation by calculation, onthe other hand, is applicable to affect-poor stimuli and is moresensitive to scope, compared with affect-rich stimuli [49].

Third, users are averse to unpleasant UX. λ > 1 specifiesthe degree of aversion to unpleasant UX, meaning users’ per-ception on those design attribute levels that are below the ref-erence point, with larger values expressing more aversion andsensitivity to unpleasant UX (see Fig. 2). The perceived UXof a product lies in a holistic impression. It is arguable that anegatively perceived product attribute alone can jeopardize thepleasant UX toward the whole product, regardless of other at-tributes that are appealing to users’ perception [35]. This effectis more pronounced for affect-rich products than for affect-poorproducts. For example, Dhar and Wertenbroch [52] reported thatparticipants were equally likely to choose a $7 music CD (anaffect-rich product) or a $7 computer disk (an affect-poor prod-uct), but were five times as likely to surrender the computer diskif they were asked to give one up.

Fourth, on the one hand, studies on social-psychological andeconomic decisions have shown that customers with positiveaffective states tend to have an optimistic bias toward pleasantUX and, therefore, take greater risks (larger α) compared withneutral individuals (see, e.g., [11] and [12]). However, positiveaffective states promote risk-averse actions for unpleasant UX(larger β) in order to sustain their positive affective states [53].On the other hand, negative affective states influence customers’risk perception in different ways. For example, based on theappraisal-tendency framework [54], compared with neutral cus-tomers, anxiety and fear are related to situations of uncertaintyand low control, which causes people to be risk-averse (smallerα and larger β), while anger is coupled with situations of cer-tainty and high control, which triggers risk-seeking behavior(larger α and smaller β) [54], [55].

2) Choice Probability: Original formulation of CPT is mo-tivated for economic outcomes (gains or losses), and thus, thechoice behavior is crafted as a subjective probability by trans-forming the objective probability of an outcome using weightingfunctions [20]. It is true that different economic outcomes occurwith varying probabilities. However, it is not the case for thechoice behavior of UX design, whereby design attribute levels

are always available for customers to choose. Therefore, mod-eling of CPT choice probabilities should be consistent with thecustomer choice behavior.

Quantitative modeling to predict choice is an established areaof research in marketing [56] and product planning [57]. Usingrandom utility discrete choice models, it is possible to predictcustomer preferences on different design attribute levels [58].The utility of a design attribute level a∗

ik to the customer isindicated by v(a∗

ik ). We can construct a closed form of choiceprobability adapted from the logit model [59]

pik = p (a∗ik ) = exp (η [v(a∗

ik )])/∑Li

k=1exp (η [v(a∗

ik )]) (2)

where η > 0 is a scaling parameter. As η → ∞, the logit be-haves like a deterministic model, while it becomes a uniformdistribution as η → 0.

3) Weighting Function: A design attribute ai with Li lev-els, i.e., A∗

i = {a∗ik}Li

, 1 ≤ k ≤ Li , can be transformed intom + n + 1 UX outcomes, where Li = m + n + 1. Arrange theoutcomes in an ascending order, i.e., v−im < · · · < vi0 < · · · <vin , or v = {vix}, with their respective choice probabilities,p−im , ..., pi0 , ..., pin , or p = {pix}. Note that vi0 correspondsto the UX of the reference level; those smaller than vi0 are re-lated to the unpleasant UX of design attribute levels; and thoselarger than vi0 are attributed to the pleasant UX of attribute lev-els. The decision maker evaluates each attribute level with theassociated choice probability, and thus the perceived UX for a∗

ik

after probability distortion can be defined as

Vik = V (a∗ik , pix) =

{vikπ+ (pix) , Δa∗

ik ≥ 0vikπ− (pix) , Δa∗

ik < 0 (3)

where π+(pix) = w+(∑n

j=x pij ) − w+(∑n

j=x+1 pij ), 0 ≤ x

≤ n − 1, π−(pix) = w−(∑x

j=−m pij ) − w−(∑x−1

j=−m pij), 1 −m ≤ x ≤ 0, π+(pin ) = w+(pin ), π−(pi,−m ) = w−(pi,−m ).

The weighting function, w, takes the following form [20]:

w(pix) = pzix

/(pz

ix + (1 − pix)z )1/z (4)

where 0 ≤ z ≤ 1 specifies the curvature of the weighting func-tion, such that z = δ stands for pleasant UX (i.e., w = w+ ),and z = θ suggests unpleasant UX (i.e., w = w−). Decreasingthe value of z makes the function become more curved. Thisfunction shows that users tend to overweigh low probabilitieswith extreme UX outcomes and underestimate moderate andhigh probabilities (see Fig. 3). One good example is that cus-tomers often overweigh the value of first-class cabins with alower choice probability, but underestimate the value of econ-omy cabins with a higher choice probability. This effect is moreevident for affect-rich products than for affect-poor products.For example, Rottenstreich and Hsee [60] showed that partici-pants were willing to pay $20 and $5 for a 1% chance to wina $500 gift certificate for a vacation to Europe (affect-rich) andschool tuition (affect-poor), respectively. However, if the chancewas 99%, participants were willing to pay $450 and $478 forthe vacation and tuition, respectively.

ZHOU et al.: PROSPECT-THEORETIC MODELING OF CUSTOMER AFFECTIVE-COGNITIVE DECISIONS 473

Fig. 3. CPT-based weighting function for the UX model.

Fig. 4. Hierarchical Bayesian parameter representation.

D. Learning Phase

In the learning phase, the parameters involved in the UXevaluation function will be estimated from the experimentaldata. Among many others, a hierarchical Bayesian method isutilized as elaborated below.

1) Hierarchical Representation of Parameters: The hierar-chical Bayesian representation of parameters involved in theUX model is shown in Fig. 4. dji = {1, 0} indicates that jthparticipant makes the ith choice about two alternative productprofiles, and if it is 1, the participant chooses Profile A and 0Profile B, where 1 ≤ i ≤ N, 1 ≤ j ≤ M . N is the total num-ber of the decisions made for each participant, and M is the totalnumber of participants. Only dji is in a shaded circle, indicat-ing observed data, while others are latent parameters (or hiddenvariables) in white circles. Whether the participant chooses Pro-file A or Profile B depends on the perceived UX calculated by(5). It can be seen that Xji is a function of individual parameters,including ηj , αj , βj , λj , δj , and θj shown in Fig. 4 with arrow“ ←.” Of all the parameters, αj , βj , δj , and j are between 0 and1. Since these parameters will have the same treatment in themodel, only the description of αj is detailed here. Accordingto Rouder and Lu [61], a probit transform model is used as fol-lows: Let Φ denote the standard normal cumulative distribution

function, and we assume that αj = Φ(zαj ), where αj ∈ [0, 1]

and zαj ∈ R. Following the probit transform model, we can have

zαj = Φ−1(αj ). Meanwhile, the probitized individual parameter

is assumed to follow an independent normal distribution at thegroup level, i.e., zα

j ∼ N(μα , (σα )2) (denoted as “ ←”), whereμα and σα are group-level parameters governing the distribu-tion of individual-level parameters (i.e., zα

j or αj ). Based onthe three-sigma rule, zα

j centers on zeros but ranges mainly in(μα − 3σα , μα + 3σα ) with a probability of 99.73%. Thus, inorder to effectively constrain the values of αj between 0 and 1,the group-level parameters also need known prior distributionson them, showing the prior knowledge about these parameters.Based on the previous research [62], for the mean, it followsa standard normal distribution, namely, μα ∼ N(0, 1). For thestandard deviation, it follows an uninformative uniform prior:σα ∼ U(0, 10). Using this manipulation, we are able to not onlyestimate individual-level parameters (e.g., αj ) to account for in-dividual differences, but also group-level parameters (e.g., μα

and σα ) to explain customer homogeneity.Similarly, parameters λj and ηj take positive values. Ac-

cording to [63], their priors follow a lognormal distribution,i.e., λi ∼ LN(μλ, (σλ)2) (ηj has the same treatment of λj ).Based on [20], the value of λj is assumed to lie in an intervalbetween 0.1 and 5, i.e., −2.30 and 1.61, on the log scale. There-fore, the mean at the group level follows a uniform distribution,i.e., μλ ∼ U(−2.30, 1.61). If an uninformative uniform priordistribution is assumed for the lognormal mean, the standard

deviation is√

1/12 (2.30 + 1.61)2 = 1.13. Hence, it is reason-able that the standard deviation at the segment level follows theuniform distribution: σλ ∼ U(0, 1.13).

Thus, Fig. 4 is summarized as follows.1) Xji ∼ f (αj , βj , λj , δj , θj , ηj ); Xji is a function of the

parameters involved in the model, based on which deci-sions can be made to generate dji . It takes either 0 or1.

2) αj , βj , λj , δj , ηj , and θj are individual-level parameters,among them αj , βj , δj , and θj are between 0 and 1,and have the same treatment. Take αj as an example,αj = Φ

(zαj

), or zα

j = Φ−1 (αj ). Thus, using the pro-bit transform, zα

j takes value from (−∞,∞), and fur-

ther zαj ∼ N

(μα , (σα )2); ηj and λj take positive values

and have the same treatment. Take λj as an example,

λj ∼ LN(μλ,

(σλ

)2), and further μλ ∼ U (−2.30, 1.61)

and σλ ∼ U (0, 1.13).3) μα , σα , μβ , σβ , μθ , σθ , μδ , σδ , μη , ση , μλ, σλ are group

level parameters, which govern individual parameter dis-tributions. They also have prior distributions that governthem, i.e., μα , μβ , μθ , μδ ∼ N (0, 1) , σα , σβ , σθ , σδ ∼U (0, 10) , μη , μλ ∼ U (−2.30, 1.61), and ση , σλ ∼U (0, 1.13).

2) Markov Chain Monte Carlo for Hierarchical BayesianEstimation: As mentioned previously, hierarchical Bayesianmodels construct the posterior distribution to estimate the pa-rameters, i.e., p(αj |dj ) ∝ p(αj )p(dj |αj ). However, it is oftenthe case that the posterior distribution is high-dimensional, com-plex, and unavailable in the closed form, and therefore, the

474 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 44, NO. 4, AUGUST 2014

analytical calculations involved are too difficult to perform [64].The emergence of Markov chain Monte Carlo (MCMC) [65] haseliminated this analytic bottleneck. The idea behind MCMC isthat it generates samples by constructing an ergodic Markovchain (i.e., irreducible and aperiodic, where irreducible meansthat every state is eventually reachable from any start, and aperi-odic means that the chain does not get caught in cycles), whichconverges after a certain number of steps to the desired pos-terior probability distribution. These samples can then be usedto approximate multidimensional integrals. Particularly, Gibbssampling is used as a special case of the Metropolis–Hastingsalgorithm [66]. The gist of Gibbs sampling is that given a mul-tivariate distribution and some initial value for each parameter,it samples each parameter from the distribution of that param-eter conditioned on the remaining parameters, making use ofthe most recent values and updating the parameter with its newvalue as soon as it has been sampled. This procedure is con-ducted recursively from the posterior conditional distributionuntil it converges. With a sufficiently large number of samples,we can compute density functions, mean values, confidence in-tervals, and other statistics of the parameters.

We define individual parameters as a 6-tuple κj = [αj , βj ,δj , θj , ηj , λj ], and κ = [κ1 , . . . , κM ] as a 6 · M tu-ple for M participants. Define π = [μα , σα , μβ , σβ , μθ ,σθ , μδ , σδ , μη , ση , μλ, σλ], containing 12 group-level parame-ters. Assuming that the parameters to estimate are denoted asϑ = [ϑ1 , . . . , ϑn ] = [κ, π], the Metropolis–Hasting sampling al-gorithm is summarized as follows [66]:

1) Initialize ϑ1 =[ϑ1

1 , . . . , ϑ1n

];

2) for i = 1, . . . , S, where S is the total number of samplesfor j = 1, . . . , n, where n is the number of parametersdraw a sample u ∼ U (0, 1);draw a candidate sample along the jth direction of theproposal distribution ϑc = q(ϑc |ϑi+1

−j , ϑij ) without

changing other coordinate values, where ϑi+1−j denotes

that all coordinate values of (i+1)th sample are givenexcept jth coordinate value;

if u < A(ϑc, ϑi

j

)= min

(1,

p(ϑc |ϑi + 1−j )q(ϑi

j |ϑc ,ϑi + 1−j )

p(ϑij |ϑ

i + 1−j )q(ϑc |ϑi

j ,ϑi + 1−j )

),

then ϑi+1j = ϑc ;

else ϑi+1j = ϑi

j .The proposal distribution is often chosen as symmetric,

such as q(ϑc |ϑi+1

−j , ϑij

)= N

(ϑi

j , (σij )

2)

so that q(ϑij |ϑc, ϑi+1

−j )and q(ϑc |ϑi

j , ϑi+1−j ) can cancel out each other. The proposal

distribution for Gibbs sampling [66] is specially chosen asq(ϑc |ϑi+1

−j , ϑij

)= p(ϑc |ϑi+1

−j ) so that the acceptance rate is al-ways 1, which leads to fast convergence.

It can be seen that the most recent draws are actually de-pendent on the previous ones. If the dependence is not strong,then convergence can happen quickly. There are different waysto assess convergence, such as diagnostic tools (see [67]). Oneinformal graphical method of assessing the degree of depen-dence is to plot the autocorrelation functions of the chains [61],that is, the correlation between values of the process of theMarkov chain of different times as a function of the time lag.One way to decrease autocorrelation is to thin the sample, using

only every kth sample. Furthermore, we run two Markov chainsto ensure convergence and comparison. For a more detaileddescription, see [68].

E. Evaluation Phase

In order to aggregate individual UX evaluation functions,considering the interdependence between individual design at-tributes, we propose to apply nested Archimedean copulas (Apopular kind of copula that allows dependence modeling in ar-bitrarily high dimensions with only one parameter, governingthe strength of dependence.) (see [46] for details). A brief de-scription is given here. The holistic perceived UX of a designprofile of I design attributes ai (1 ≤ i ≤ I) with its ki–th levelis aggregated by an Archimedean copula structure, i.e.,

X(a,p) = C(x1k1 , ..., xIkI)

= c1ϕ−1

[ ∏I

i=1ϕ(li + (1 − li)xiki

)]

+ c2 (5)

where xikiis between 0 and 1 by normalizing Vik in (3), indi-

cating UX of individual-attribute value and X is the aggregatedmultiattribute value; c1 = 1/(1 − ϕ−1 [

∏Ii=1 ϕ(li)]), c2 = 1 −

c1 , 0 ≤ li < 1, and the generating function ϕ is 1) continuouson the domain xki

∈ [0, 1]; 2) strictly increasing on the do-main [0, 1]; and 3) ϕ(0) = 0 and ϕ(1) = 1. It has the formϕ(li) = (1 − exp(−ςli))/(1 − exp(−ς)), where ς ∈ R\{0}.

The perceived UX based on CPT involves various mentalprocesses that are thought to drive decision making to opti-mize UX [63]. According to (5), under CPT, a UX prospectX(a1 ,p1) is preferred to another prospect X(a2 ,p2) for aspecific customer if and only if X(a1 ,p1) � X(a2 ,p2) andis indifferent to each other if X(a1 ,p1) ∼ X(a2 ,p2). Basedon the descriptions above, preferences in terms of UX are de-termined jointly by a subjective value function that evaluatesindividual UX of specific design attribute level with regard toa reference point, and by the decision weights that capture anindividual’s distortion of choice probability. Furthermore, theshaping parameters embedded in the subjective value functioncapture users’ physiological processes, including risk attitudes,cognitive tendencies, and affective influences. Finally, a holis-tic measure of multiattribute UX is obtained by aggregatingindividual-attribute UX. Under the circumstances of product de-sign for multiple design attributes A = {ai}I and each with sev-eral levels A∗

i = {a∗ik}I×Li

, the cumulative prospect-theoreticUX model can be used to evaluate customers’ holistic UX ofalternative design profiles in the design space.

IV. EMPIRICAL STUDY FOR MODEL PARAMETER

ESTIMATION AND VALIDATION

A. Background

The experiment focused on the aircraft cabin interiordesign and aimed to create positive UX in the aircraft cabin.In order to disguise proprietary information, the aircraft cabininterior design was simplified deliberately. The design attributes

ZHOU et al.: PROSPECT-THEORETIC MODELING OF CUSTOMER AFFECTIVE-COGNITIVE DECISIONS 475

TABLE IDESIGN ATTRIBUTES AND LEVELS OF AIRCRAFT CABIN INTERIOR DESIGN

and their attribute levels are shown in Table I. For example,one important aspect that influences cabin UX is the personalspace for passengers, including legroom and seat aircraft space.Legroom is not just a matter of comfort—it is a matter of health,too. Seating space on airplanes is approximated by width andseat pitch—the distance from a point in the headrest to the samepoint in the headrest in front. For another example, a sculptedceiling with gentle curves makes an airplane cabin feel morespacious. Curved shapes are friendlier and are perceived to besafer than hard edges.

B. Cabin Profiles

Given all the design attributes and their levels in Table I,a total number of 5 × 2 × 38 = 65610 combinations can beconstructed. To overcome such an explosion of configurations,orthogonal product profiles were used. Twenty-seven orthogonalproduct profiles were generated based on the principle of designof optimal experiment [69] with SPSS 15.0 in Table II, in whichcolumns 2–11 indicate the specification of the product profiles.

C. Hypothesis

Based on the discussion in Section III, we summarized fourhypotheses addressing the main affective and cognitive influ-ences on choice decision making in the CPT-based UX model:

H1: When the participant is joyful and excited (positive af-fect), the value function would have a larger α and a smaller βthan when the participant is neutral (no particular affect).

H2: When the participant is anxious (negative affect), thevalue function would have a smaller α and a larger β than whenthe participant is neutral.

H3: The value function would have a smaller α, a smallerβ, and a larger λ, and the weighting function would have asmaller z (δ or θ) of affect-rich products than those of affect-poor products.

H4: The value of λ would be larger than 1, indicating aversionto unpleasant UX.

D. Emotion Elicitation

Self-elicited methods by imagination or imagery (see, e.g.,[70] and [71]) were used in this study to elicit target affectivestates from the participants. It required participants to be in-volved in the target affective states deliberately by recalling orimagining a certain situation. In order to facilitate the proce-dure, two descriptions related to cabin interior design were pro-vided. First, it was assumed that the aircraft cabin designed wasemployed on a trip to Paris. For joy and excitement, positivedescriptions and corresponding images about Paris were pro-vided: Paris is the world’s leading tourism destination. AmongParis’ first mass attractions drawing international interests arethe Eiffel Tower, the world’s most-visited art museum, the Lou-vre, housing many works of art, including the Mona Lisa andthe Venus de Milo statue. . .. This description projected the ex-pected emotions positively so that participants would have joy,happiness, and excitement about the trip as a result. For anxiety,questions were shown to the participants: How will I deal withthe local language? How expensive will things be? Will I havegood weather? Will my bank hold my credit card? Do I forget topack something? These questions typically evoked anxiety sincethey described situations with unpredictable or uncontrollableevents [72]. Finally, neutral was elicited with no descriptionbut with a white blank picture. Neutral was used as a baselinelevel and a control condition for comparison. Furthermore, onlythose whose self-reported affective states were consistent withthe target ones were used in the experiment.

E. Participants

University students aged between 20 and 30 with genderbalance were recruited. Among them, 60 participated in Study1 and were divided into three groups averagely with genderbalance. Participants in each group were asked to elicit oneof the target affective states (i.e., joy and excitement, neutral,or anxiety), respectively, and then asked to evaluate differentdesign profiles of the aircraft cabin. In such a way, the group ofparticipants in a neutral state was regarded as a control group andthe other two as treatment groups. Forty participated in Study 2

476 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 44, NO. 4, AUGUST 2014

TABLE IIAIRCRAFT CABIN INTERIOR DESIGN PROFILES FOR EVALUATIONS

and were divided into two groups averagely with gender balance.Participants in each group were asked to evaluate one kind ofaircraft cabin interior designs, i.e., affect-rich or affect-poor,respectively. These two groups were regarded as two differenttreatments for comparison.

F. Procedure

Study 1: After the participants had signed the consent forms,they were told that they were going to take a round- trip flightfrom Atlanta, USA, to Paris, France, and they were asked to self-report cabin UX for different cabin profiles. They were thenbriefed on the procedure of the study. First, they were askedto navigate a virtual aircraft cabin, focusing on the design at-tributes. The immersed navigation was operated with a keyboardby the participant who was seated comfortably in an armchair.Although the cabin environment they navigated was not entirelythe same as the design profiles offered, it helped them becomefamiliar with the overall environment and improve the accuracyof self-reported UX. Second, they were asked to read the de-scriptions of the travel (including the associated images) and beinvolved in the target affective states deliberately by imaginingthe situation as described. Third, each participant was showna series of concepts as shown in Table II and was required toself-report UX with regard to different design attribute levelsand make decisions between two alternative product profiles.

Study 2: All the participants were briefed that they would takea flight for a round trip from Atlanta, USA, to Paris, France, andthey were required to self-report the cabin UX. However, halfof the participants were told that they were on a vacation andbeautiful pictures of Paris were shown to them; the rest weretold that they were on business with a tight schedule. Otherprocedures were the same as Study 1 except that they were notasked to elicit emotions. Instead, the first group of participantswas considered to self-report the aircraft cabin UX associatedwith an affect-rich vacation, whereas the second group wasassociated with an affect-poor trip.

G. Data Collection

The UX was measured on a scale between −10 and 10, where10 indicated extremely pleasant and −10 extremely unpleasantwith regard to individual design attribute levels [see Fig. 5(a)].Further, they were required to make decisions between two

Fig. 5. Data collection. (a) Self-reported UX of individual design attributes(only part shown here). (b) Decision making between alternative options.

alternative design profiles, as shown in Fig. 5(b). Of all the 27design profiles (in random order for different participants), eachparticipant was required to make 26 decisions (e.g., Profile 1versus Profile 2, if Profile 2 was preferred, the next comparisonwas Profile 2 versus Profile 3, and so on. It thus resulted in 26choices). They were preprocessed and structured for analysisof hierarchical Bayesian model to estimate the parameters fordifferent groups of participants.

Based on the experiment of two studies, the datasets for dif-ferent purposes of the studies were produced. For Study 1,three datasets were generated for three different target emo-tions, i.e., joyful and excited, anxious, and neutral. For Study2, two datasets were generated for two types of products, i.e.,affect-poor and affect-rich. For both studies, the decision datawere divided into a training dataset (80% of the data) for param-eter estimation and a test dataset (20% of the data) for modelvalidation. This process was run three times to generate the av-eraged results. Two Markov chains were generated. Each chainhad 20000 samples with 5000 burn-in samples, and only ev-ery tenth sample was collected. Therefore, for each chain, 1500samples were valid to estimate the posterior distributions ofthe parameters. Convergence was confirmed by autocorrelationgraphs and visual inspection [61], [63].

ZHOU et al.: PROSPECT-THEORETIC MODELING OF CUSTOMER AFFECTIVE-COGNITIVE DECISIONS 477

TABLE IIIRESULTS OF PARAMETER ESTIMATION IN THREE AFFECTIVE STATES

TABLE IVCLASSIFICATION BASED ON CANONICAL DISCRIMINANT ANALYSIS

FOR THREE AFFECTIVE STATES

V. RESULTS AND VALIDATION

A. Affective Influence

The parameters in the UX model for three different affectivestates were estimated in Table III, and their posterior densityfunctions were estimated using the ksdensity function in Matlabas illustrated in Fig. 6 (except η). In order to test H1 and H2,single factor ANOVA was used. It showed that there were sig-nificant differences among three affective states (α: F (2,57) =19.76, p < 0.001; β: F (2,57) = 6.02, p < 0.01). Bonferronipost-hoc analysis showed that α in anxiety is significantlysmaller than those in the other two affective states (p < 0.05);β in joy and excitement is significantly smaller than those inneutral and anxiety (p < 0.01). The results showed that anxiousparticipants tend to be more risk-averse than neutral ones forpleasant UX, and joyful, excited participants for both pleasantand unpleasant UX; joyful and excited participants tend to bemore risk-seeking than neutral and anxious participants for bothunpleasant and pleasant UX. These results supported H1 and H2(partially). λ values in three affective states were significantlylarger than 1, indicating that all the participants were averse tounpleasant UX. Therefore, H4 was accepted.

Moreover, for the individual parameters, we conducted acanonical discriminant analysis [75], in which discriminantfunctions were linear combinations of individual parameters andwere used to predict participants into different affective groups.Fig. 7(a) shows the scatter plot of the individual parametersmapped in two canonical discriminant functions. It can be seenthat people in joy and excitement and in anxiety seem separatedfrom each other, while people in neutral states seem scatteredwidely into other two groups. Table IV shows the classificationresults in terms of recall, precision, and F1 measure [27]. Pre-cision shows how well the model predicts (i.e., a measure ofexactness), and recall accounts for how well the model does not

Fig. 6. Posterior probability density functions for three different affectivestates: (a) α and β , (b) θ and δ, and (c) λ.

miss the target (i.e., a measure of completeness). In Table IV,precision and recall are 75.0% and 90.0% for joy and excite-ment, 66.7% and 60.0% for neutral, and 88.2% and 80.0% foranxiety, respectively. The F1 measure combines the precisionand the recall and is the harmonic mean of them. It thus gives theoptimal accuracy. The mean F1 measure is 76.3% in Table IV.

478 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 44, NO. 4, AUGUST 2014

Fig. 7. Canonical discriminant analysis for (a) three affective groups and(b) two types of products.

TABLE VRESULTS OF PARAMETER ESTIMATION FOR TWO TYPES OF PRODUCTS

B. Affect-Rich Versus Affect-Poor Products

As mentioned previously, two types of products are identi-fied, i.e., affect-poor and affect-rich. The parameters involvedin the UX model for these two types were estimated in Table V,and their posterior probability density functions were estimatedusing ksdensity function in MATLAB, as illustrated in Fig. 8(except η). Using single factor ANOVA, it is found that both αand β were significantly smaller (α: F (1,38) = 9.31, p < 0.01;β: F (1,38) = 8.26, p < 0.01) in the affect-rich trip than thosein the affect-poor trip. This showed that the assumptions ofvaluation by feeling in affect-rich products and valuation bycalculation in affect-poor products are accepted. However, nosignificant difference was found between the values of λ in two

Fig. 8. Posterior probability density functions in two different product types:(a) α and β , (b) θ and δ, and (c) λ.

types of products (λ: F (1,38) = 2.65, p = 0.11). This showedthat participants were equally sensitive to two types of productsfor unpleasant UX. The values of δ and θ were both significantlysmaller in the affect-rich trip than those in the affect-poor trip (δ:F (1,38) = 7.58 p < 0.01; θ: F (1,38) = 8.65, p < 0.01). Thissupported (see Fig. 2) that the weighting function was morecurved in the affect-rich trip than that in the affect-poor one.Therefore, H3 is supported except the value of λ. However, λ

in two different types of trips were significantly larger than 1,

ZHOU et al.: PROSPECT-THEORETIC MODELING OF CUSTOMER AFFECTIVE-COGNITIVE DECISIONS 479

TABLE VICLASSIFICATION BASED ON CANONICAL DISCRIMINANT ANALYSIS

FOR TWO TYPES OF PRODUCTS

TABLE VIIDECISION PREDICTION ACCURACY OF DIFFERENT CPT MODEL

indicating all the participants were averse to unpleasant UX.Hence, H4 was accepted.

Similarly, we also conducted a canonical discriminant analy-sis for two types of products (only one canonical discriminantfunction was produced). Fig. 7(b) shows the histogram of thecanonical discriminant function in which affect-rich productsprimarily have positive values while affect-poor products mainlyhave negative values. Table VI shows the classification resultsbased on the canonical discriminant analysis. The precision andrecall are 73.9% and 85.0% for affect-poor products and 82.4%and 70.0% for affect-rich products, respectively. The averageF1 measure is 77.4%.

Besides, we acknowledged the amount of uncertainty asso-ciated with each parameter by their posterior distributions, asshown in Figs. 6 and 8. Furthermore, substantial uncertaintyin drawing inferences about the UX was also indicated by theparameter estimated for individual participants, showing eachone had a different UX evaluation function. For example, themean value of the parameters of participant 8 were α = 0.47,β = 0.76, λ = 2.83, δ = 0.27, θ = 0.24, and η = 2.15, and ofparticipant 16 were α = 0.39, β = 0.68, λ = 2.31, δ = 0.30,θ = 0.24, and η = 1.98. The values were unique for a particularparticipant, and thus, product customization can be realized byapplying these parameters for UX evaluation and prediction.

C. Prediction Accuracy and Optimal Cabin Profile

In order to validate the proposed model, decision-makingprediction with a test dataset was conducted for three times. Theprediction accuracies for different situations were summarizedin Table VII for five different models. The prediction accuracywas based on the decision making between the two alternativedesign profiles [see Fig. 5(b)] according to (5). It is defined asthe number of accurately predicted decisions divided by the totalnumber of decisions made.

Furthermore, we obtained the UX prediction functions for allthe design profiles in the design space. For example, for the de-sign profiles 1 and 3 in Table II, the evaluations (see Table VIII)by participants in the neutral group were calculated using (1)–(5) with the estimated parameters. The values of aggregated UXfor profiles 26 and 27 are 0.10 and 0.65, respectively. There-fore, between the two design options, profile 27 was preferred.Moreover, for the different participant groups, the aggregatedUX was shown in Fig. 9. It showed that the basic trend wasconsistent among five participant groups and the optimal cabinprofiles were profile 14 (0.89 for anxiety), profile 14 (0.92 forjoy and excitement), profile 14 (0.93 for neutral), profile 9 (0.79affect-poor), and profile 14 (0.86 affect-rich), respectively.

VI. DISCUSSIONS

A. Affective Factors in Decision Making for the UserExperience Design

Based on the individual parameters, we find that systematicaldifferences exist among the parameters in the subjective valuefunction for participants in three different affective states (seeTable III). Consistent with this finding, the canonical discrimi-nant analysis also predicted three affective groups. Specifically,we found that anxious participants tend to be more risk-aversethan neutral ones for pleasant UX, and joyful, excited partic-ipants for both pleasant and unpleasant UX; and that joyfuland excited participants tend to be more risk-seeking than neu-tral and anxious participants for both unpleasant and pleasantUX. We also identified the differences of six individual pa-rameters with regard to affect-rich and affect-poor products, asevidenced in Tables V and VI. Specifically, participants tend tovalue affect-rich products by feeling and value affect-poor prod-ucts by calculation. Furthermore, the results were also validatedby decision-making prediction based on the parameters esti-mated. We controlled the experiment setting strictly the same,except that the participants were in different affective states orproduct types. In this sense, the differences of the parametersamong different groups are mainly attributed to the affectivefactors rather than cognitive factors. F1 measures for the threeaffective groups and two types of products are both around75%, and quite some overlaps of parameter distributions wereobserved in Figs. 6 and 8. It is possible that the hint of affectelicitation in the experimental setting is so strong that the differ-ences were not that noticeable from the posterior distributionsof the parameters. This is also consistent with the results ofFig. 9. However, in other words, it can be interpreted as thesubstantial uncertainty involved in the posterior distributions ofthe parameters. This is, otherwise, not possible for point esti-mate in maximum likelihood estimation, which often leads toover-confident predictions. As a common practice, the resultswe obtained are based on the posterior means, which often showthe group homogeneity. The hierarchical Bayesian model allowsus to calculate parameters for individuals that show individualdifferences. It inherently implies a personalization strategy forindividual customers, which compels the producer to examinedifferent combinations of existing design attributes and valueprofiles to anticipate and adapt to customers’ latent needs. Such

480 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 44, NO. 4, AUGUST 2014

TABLE VIIIUX COMPARISON BETWEEN DESIGN PROFILE 26 AND PROFILE 27

Fig. 9. Aggregated UX value for different cabin profiles of five groups.

a strategy of product innovation is more likely to guaranteepleasurable UX at both the individual and group levels.

B. Implications for the User Experience Design

Based on the UX evaluation function, the aggregated per-ceived UX can be predicted for a particular design profile, whichis a key performance indicator for evaluation of alternative de-sign profiles. Meanwhile, the contribution of individual designattributes to pleasant or unpleasant UX can also be specified.This provides sensitivity analysis for the value-added UX de-sign. The results partially support H2 and H3 and support H1 andH4. It has several implications for the UX design as explainedbelow: First, for three different affective states, the parame-ters involved in the value function are significantly different.This implies that more investigation needs to be conducted toavoid “affective bias” from customers’ subjective perceptionand recognize the design’s actual contribution to UX fulfill-ment. Second, affect-rich products usually have a smaller valueof α and a larger value of β. It means that the absolute magni-tude of pleasant UX increases at a slower rate than affect-poorproducts, and vice versa for unpleasant UX. However, it will

have a larger value of λ, indicating that if one design attributeleads to unpleasant UX, the holistic UX toward the productwill be severely affected. Therefore, more attention should bepaid so that no attributes will lead to negative UX. Third, howthe product is framed (i.e., affect-rich or affect-poor) also in-fluences customers’ UX perception and evaluation. Therefore,companies should be aware of the fact that products with highaffective quality (the ability to elicit positive emotions) but withlow probabilities are perceived to have superior UX. For ex-ample, for aircraft cabin interior design, first-class cabins withsuperior design attributes, including spacious personal space,better food and beverage, and booking and checking in priority,etc., are highly deemed in terms of perceived UX.

C. Limitations

As an exploratory study, the approach proposed in this re-search suffers a few limitations. First, not all the hypothesesare supported. It may be due to the affect elicitation techniquesthat the flight is hypothetical and the description is plain with-out vivid multimedia so that the elicited affective states couldbe different from the target affective states or the intensity is

ZHOU et al.: PROSPECT-THEORETIC MODELING OF CUSTOMER AFFECTIVE-COGNITIVE DECISIONS 481

not strong enough. Other possible techniques, like videos, maybe more effective in future studies. Furthermore, the designprofiles are mainly based on the word description, which can-not mimic the real aircraft cabin environment and thus lacksecological validity. In order to improve ecological validity, vir-tual reality can help create the cabin environment with lowercosts in a shorter time compared with real aircraft cabin proto-types. Second, we only consider three different affective statesfor a small sample size. It is cautious to generalize the resultsto other affective states. Third, there are numerous empiricalassumptions made in the hierarchical Bayesian model for pa-rameter estimation, and thus, the method proposed here maynot well accommodate the link between core cognitive decisionmaking and affective influence in terms of parameter shaping.For example, a large amount of subjectivity is involved in se-lecting prior distributions. In many situations, we believe thatwe can prudently select proper prior distributions to effectivelymake use of the prior knowledge in the estimation process, de-spite the fact that we can use uninformative priors when noparticular prior knowledge is available. In this paper, we choosethe normal distribution for the group-level parameters between0 and 1, and the lognormal distribution for the rest based onprevious research [63]. We then select either normal or fixeduninformative priors (i.e., uniform distributions) for their pri-ors. These priors actually control the group-level parameters incertain ranges, which provide a very efficient method for esti-mation since it can fully use the knowledge of the distributionalstructure [73]. The efficacy is validated by the results. However,it does not exclude the possibility of better priors. For example,the Beta distribution is often assumed for parameters between 0and 1 (see, e.g., [74]).

VII. CONCLUSION

UX has become the key success factor in the product designand models that can quantify and evaluate UX have receivedmuch attention. This research proposes a CPT-based affective-cognitive decision-making model for the quantification, predic-tion, and evaluation of UX. The model includes four integratedphases, i.e., the perceptual phase, the affective-cognitive rea-soning phase, the learning phase, and the evaluation phase. Itaddresses the fundamental issues involved in affective-cognitivedecision making for UX design, including affective and cogni-tive influences on the CPT-based value function and the weight-ing function. A hierarchical Bayesian model is utilized to esti-mate the parameters involved in UX evaluation function, con-sidering the inherent uncertainty by incorporating posterior den-sity functions of the estimated parameters. An empirical studyis conducted to test the hypotheses proposed. Results show thataffective states influence decision making involved in the UXdesign. The model is also able to predict UX decision makingwith moderate accuracies and the optimal product profile canbe identified with the UX evaluation function. We believe thatthis new model provides the flexibility and comprehensivenessneeded to explain decision-making behavior involved in the UXdesign.

REFERENCES

[1] C. K. Cho, Y. S. Kim, and W. J. Lee, “Economical, ecological and expe-rience values for product-service systems,” presented at the 7th DesignEmotion Conf., Chicago, IL, USA, 2010.

[2] R. W. Picard, Affective Computing. Cambridge, MA, USA: MIT Press,1997.

[3] J. G. Falcioni, “Make it work,” ASME Mech. Eng. Mag., vol. 130, no. 2,p. 4, Feb. 2008.

[4] A. S. Brown, “The new point of view: Focus on design for human factors,”ASME Mech. Eng. Mag., vol. 130, no. 2, pp. 22–27, Feb. 2008.

[5] C. D. Wickens and J. G. Hollands, Engineering Psychology and HumanPerformance, 3rd ed. Englewood Cliffs, NJ, USA: Prentice-Hall, 1999.

[6] D. Kahneman, “A perspective on judgment and choice: Mapping boundedrationality,” Amer. Psychologist, vol. 58, no. 9, pp. 697–720, 2003.

[7] H. Simon, “Affect and cognition: Comments,” in Affect and Cognition:The 17th Annual Carnegie Symposium on Cognition, M. S. Clark andS. T. Fiske, Eds. Mahwah, NJ, USA: Erlbaum, 1982, pp. 333–342.

[8] U. Khan, R. Dhar, and K. Wertenbroch, (2004). A behavioral decision the-oretic perspective on hedonic and utilitarian choice. [Online]. Available:www.insead.edu/facultyresearch/research/doc.cfm?did = 1411

[9] D. W. Rook, “The buying impulse,” J. Consumer Res., vol. 14, no. 2,pp. 189–199, 1987.

[10] P. Slovic, M. L. Finucane, E. Peters, and D. G. Macgregor, “Risk as anal-ysis and risk as feelings: Some thoughts about affect, reason, risk, andrationality,” Risk Anal., vol. 24, no. 2, pp. 311–322, 2004.

[11] A. Bracha and D. Brown, “Affective decision making: A theory of opti-mism bias,” Games Econ. Behav., vol. 75, no. 1, pp. 67–80, 2012.

[12] H. Ahn, “Modeling and analysis of affective influences on human ex-perience, prediction, decision making, and behavior,” Ph.D. dissertation,Massachusetts Inst. Technol., Cambridge, MA, USA, 2010.

[13] D. Kahneman, “Experienced utility and objective happiness: A moment-based approach,” in Choices, Values, and Frames, D. Kahneman andA. Tversky, Eds. New York, NY, USA: Cambridge Univ. Press, 2000,pp. 673–692.

[14] E. Brandstatter, G. Gigerenzer, and R. Hertwig, “The priority heuristic:Making choices without trade-offs,” Psychological Rev., vol. 113, no. 2,pp. 409–432, 2006.

[15] H. Ahn and R. W. Picard, “Affective-cognitive learning and decision mak-ing: A motivational reward framework for affective agents,” presented atthe 1st Int. Conf. Affective Comput. Intell. Interaction, Beijing, China,2005.

[16] K. R. Scherer, A. Shorr, and T. Johnstone, Appraisal Processes in Emo-tion: Theory, Methods, Research. New York, NY, USA: Oxford Univ.Press, 2001.

[17] J. Gratch and S. Marsella, “A domain-independent framework for model-ing emotion,” Cognitive Syst. Res., vol. 5, no. 4, pp. 269–306, 2004.

[18] R. P. Marinier, III, J. E. Laird, and R. L. Lewis, “A computational unifi-cation of cognitive behavior and emotion,” Cognitive Syst. Res., vol. 10,no. 1, pp. 48–69, 2009.

[19] D. Kahneman and A. Tversky, “Prospect theory: An analysis of decisionunder risk,” Econometrica, vol. 62, no. 6, pp. 1251–1289, 1979.

[20] A. Tversky and D. Kahneman, “Advances in prospect theory: Cumulativerepresentation of uncertainty,” J. Risk Uncertainty, vol. 5, no. 4, pp. 297–323, 1992.

[21] G. W. Harrison and E. E. Rutstrom, “Expected utility theory and prospecttheory: One wedding and a decent funeral,” Exp. Econ., vol. 12, no. 2,pp. 133–158, 2009.

[22] M. D. Lee and B. R. Newell, “Using hierarchical bayesian methods toexamine the tools of decision-making,” Judgment Decision Making, vol. 6,no. 8, pp. 832–842, 2011.

[23] M. Steyvers, M. D. Lee, and E. J. Wagenmakers, “A bayesian analysis ofhuman decision-making on bandit problems,” J. Math. Psychol., vol. 53,no. 3, pp. 168–179, 2009.

[24] E. L. C. Law and P. Van Schaik, “Modelling user experience—An agendafor research and practice,” Interac. Comput., vol. 22, no. 5, pp. 313–322,2010.

[25] M. Hassenzahl, “The interplay of beauty, goodness and usability in inter-active products,” Human Comput. Interact., vol. 19, no. 4, pp. 319–349,2004.

[26] P. van Schaik and J. Ling, “Modelling user experience with web sites:Usability, hedonic value, beauty and goodness,” Interact. Comput., vol. 20,no. 3, pp. 419–432, 2008.

[27] F. Zhou, X. Qu, M. G. Helander, and J. Jiao, “Affect prediction fromphysiological measures via visual stimuli,” Int. J. Human-Comput. Stud.,vol. 69, no. 12, pp. 801–819, 2011.

482 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 44, NO. 4, AUGUST 2014

[28] T. Partala, V. Surakka, and T. Vanhala, “Real-time estimation of emotionalexperiences from facial expressions,” Interact. Comput., vol. 18, no. 2,pp. 208–226, 2006.

[29] D. Cyr, M. Head, and A. Ivanov, “Design aesthetics leading to m-loyaltyin mobile commerce,” Inform. Manage., vol. 43, no. 8, pp. 950–963,2006.

[30] P. Zhang, N. Li, and H. Sun, “Affective quality and cognitive absorption:Extending technology acceptance research,” presented at the 39th Annu.Hawaii Int. Conf. Syst. Sci., Kauai, HI, USA, 2006.

[31] T. Partala and A. Kallinen, “Understanding the most satisfying and un-satisfying user experiences: Emotions, psychological needs, and context,”Interact. Comput., vol. 24, no. 1, pp. 25–34, 2012.

[32] F. Zhou, R. J. Jiao, Q. Xu, and K. Takahashi, “User experience modelingand simulation for product ecosystem design based on Fuzzy reasoningpetri nets,” IEEE Trans. Syst., Man Cybern. A, Syst. Humans, vol. 42,no. 1, pp. 201–212, Jan. 2012.

[33] J. Park, S. H. Han, H. K. Kim, S. Oh, and H. Moon, “Modeling userexperience: A case study on a mobile device,” Int. J. Ind. Ergonom.,vol. 43, pp. 187–196, 2013.

[34] J. Kim and S. H. Han, “A methodology for developing a usability indexof consumer electronic products,” Int. J. Ind. Ergonom., vol. 38, no. 3/4,pp. 333–345, 2008.

[35] F. Zhou, J. R. Jiao, D. Schaefer, and S. Chen, “Hybrid association min-ing and refinement for affective mapping in emotional design,” ASME J.Comput. Inform. Sci. Eng., vol. 10, no. 3, pp. 031010–031019, 2010.

[36] H. Zhang and S. Liu, “Design of autonomous navigation system basedon affective cognitive learning and decision-making,” presented at the Int.Conf. Robot. Biomimetics, Guilin, China, 2009.

[37] T. E. Power, L. C. Swartzman, and J. W. Robinson, “Cognitive-emotionaldecision making (CEDM): A framework of patient medical decision mak-ing,” Patient Educ. Counseling, vol. 83, no. 2, pp. 163–169, 2011.

[38] B. Penolazzi, P. Gremigni, and P. M. Russo, “Impulsivity and rewardsensitivity differentially influence affective and deliberative risky decisionmaking,” Personality Individual Differences, vol. 53, no. 5, pp. 655–659,2012.

[39] A. Bechara, H. Damasio, D. Tranel, and A. Damasio, “Deciding advan-tageously before knowing the advantageous strategy,” Sci., New Series,vol. 275, no. 5304, pp. 1293–1295, 1997.

[40] G. A. Hazelrigg, “A framework for decision-based engineering design,”J. Mech. Design, vol. 120, no. 4, pp. 653–658, 1998.

[41] W. Chen, K. E. Lewis, and L. C. Schmidt, “The open workshop ondecision-based design,” in Decision Making in Engineering Design,K. E. Lewis, W. Chen, and L. C. Schmidt, Eds. New York, NY, USA:ASME, 2006.

[42] D. L. Thurston, “Utility function fundamentals,” in Decision Makingin Engineering Design, K. E. Lewis, W. Chen, and L. C. Schmidt, Eds.New York, NY, USA: ASME, 2006, pp. 15–19.

[43] S. Orsborn, J. Cagan, and P. Boatwright, “Quantifying aesthetic form pref-erence in a utility function,” J. Mech. Design, vol. 131, no. 6, pp. 061001-1–061001-10, 2009.

[44] H. Xu, J. Zhou, and W. Xu, “A decision-making rule for modeling travel-ers’ route choice behavior based on cumulative prospect theory,” Transp.Res. C, Emerging Technol., vol. 19, no. 2, pp. 218–228, 2011.

[45] J. Y. J. Chow, G. Lee, and I. Yang, “Genetic algorithm to estimate cumu-lative prospect theory parameters for selection of high-occupancy-vehiclelane,” Transp. Res. Rec., J. Transp. Res. Board, vol. 2, no. 2157, pp. 71–77,2010.

[46] F. Zhou and R. J. Jiao, “A nested multivariate utility copulas approach toaggregating user experience partworths for aircraft cabin interior design,”presented at the ASME Int. Design Eng. Tech. Conf. Comput. Inform.Eng. Conf., Portland, OR, USA, 2013.

[47] M. Hassenzahl and N. Tractinsky, “User experience—A research agenda,”Behav. Inform. Technol., vol. 25, no. 2, pp. 91–97, 2006.

[48] P. C. Ellsworth and K. R. Scherer, “Appraisal processes in emotion,”in Handbook of Affective Sciences, R. J. Davidson, K. R. Scherer, andH. H. Goldsmith, Eds. New York, NY, USA: Oxford Univ. Press, 2003,pp. 572–595.

[49] C. K. Hsee and Y. Rottenstreich, “Music, pandas, and muggers: On theaffective psychology of value,” J. Exp. Psychol.: Gen., vol. 133, no. 1,pp. 23–30, 2004.

[50] W. H. Desvousges, F. Johnson, R. Dunford, S. Hudson, K. Wilson, andK. Boyle, “Measuring resource damages with contingent valuation: Tests

of validity and reliability,” in Contingent Valuation: A Critical Assessment,J. A. Hausman, Ed. Amsterdam, The Netherlands: North Holland, 1993,pp. 91–164.

[51] D. Kahneman, I. Ritov, and D. Schkade, “Economic preferences or attitudeexpressions? An analysis of dollar responses to public issues,” J. RiskUncertainty, vol. 19, no. 1–3, pp. 203–235, 1999.

[52] R. Dhar and K. Wertenbroch, “Consumer choice between hedonic andutilitarian goods,” J. Marketing Res., vol. 37, no. 1, pp. 60–71, 2000.

[53] A. M. Isen, “An influence of positive affect on decision making in complexsituations: Theoretical issues with practical implications,” J. Consum.Psychol., vol. 11, no. 2, pp. 75–85, 2001.

[54] J. S. Lerner and D. Keltner, “Fear, anger, and risk,” J. Personality SocialPsychol., vol. 81, no. 1, pp. 146–159, 2001.

[55] E. Gambetti and F. Giusberti, “The effect of anger and anxiety traits oninvestment decisions,” J. Econ. Psychol., vol. 33, no. 6, pp. 1059–1069,2012.

[56] J. J. Louviere, D. A. Hensher, and J. D. Swait, Stated Choice Methods:Analysis and Applications. New York, NY, USA: Cambridge Univ.Press, 2000.

[57] K. Lewis, W. Chen, and L. Schmidt, Eds., Decision Making in Engineer-ing Design. New York, NY, USA: ASME, 2006.

[58] P. E. Green and A. M. Krieger, “Models and heuristics for product lineselection,” Marketing Sci., vol. 4, no. 1, pp. 1–19, 1985.

[59] K. E. Train, Discrete Choice Methods with Simulation. Cambridge,U.K.: Cambridge Univ. Press, 2003.

[60] Y. Rottenstreich and C. K. Hsee, “Money, kisses, and electric shocks:On the affective psychology of risk,” Psychological Sci., vol. 12, no. 3,pp. 185–190, 2001.

[61] J. N. Rouder and J. Lu, “An introduction to bayesian hierarchical modelswith an application in the theory of signal detection,” Psychonomic Bull.Rev., vol. 12, no. 4, pp. 573–604, 2005.

[62] A. Gelman and J. Hill, Data Analysis Using Regression and Multi-level/Hierarchical Models. Cambridge, U.K.: Cambridge Univ. Press,2007.

[63] H. Nilsson, J. Rieskamp, and E. J. Wagenmakers, “Hierarchical Bayesianparameter estimation for cumulative prospect theory,” J. Math. Psychol.,vol. 55, no. 1, pp. 84–93, 2011.

[64] D. J. Lunn, A. Thomas, N. Best, and D. Spiegelhalter, “WINBugs—Abayesian modelling framework: Concepts, structure, and extensibility,”Statist. Comput., vol. 10, no. 4, pp. 325–337, 2000.

[65] R. M. Neal, “Probabilistic inference using Markov chain Monte Carlomethods,” Univ. Toronto, Toronto, ON, Canada, Tech. Rep. CRG-TR-93-1, 1993.

[66] S. Geman and D. Geman, “Stochastic relaxation, Gibbs distributions andthe Bayesian restoration of images,” IEEE Trans. Pattern Anal. Mach.Intell., vol. PAMI-12, no. 6, pp. 609–628, Nov. 1984.

[67] M. K. Cowles and B. P. Carlin, “Markov chain Monte Carlo convergencediagnostics: A comparative review,” J. Amer. Statist. Assoc., vol. 91,no. 434, pp. 883–904, 1996.

[68] F. Zhou and R. J. Jiao, “Hierarchical Bayesian parameter estimation formodeling and analysis of user affective influence,” presented at the ASMEInt. Design Eng. Tech. Conf. Comput. Inform. Eng. Conf., Portland, OR,USA, 2013.

[69] S. K. Nair, L. S. Thakur, and K. Wen, “Near optimal solutions for productline design and selection: Beam search heuristics,” Manage. Sci., vol. 41,no. 5, pp. 767–785, 1995.

[70] R. Sinha and O. Parsons, “Multivariate response patterning of fear andanger,” Cognition Emotion, vol. 10, no. 2, pp. 173–198, 1996.

[71] R. W. Picard, E. Vyzas, and J. Healey, “Toward machine emotional intel-ligence: Analysis of affective physiological state,” IEEE Trans. PatternAnal. Mach. Intell., vol. 23, no. 10, pp. 1175–1191, Oct. 2001.

[72] M. E. P. Seligman, Helplessness: on Depression, Development, andDeath. San Francisco, CA, USA: Freeman, 1975.

[73] H. T. Banks, S. Grove, S. Hu, and Y. Ma, “A hierarchical Bayesianapproach for parameter estimation in HIV models,” Inverse Problems,vol. 21, no. 6, pp. 1803–1822, 2005.

[74] N. Bouguila and T. Elguebaly, “A fully bayesian model based on re-versible jump mcmc and finite beta mixtures for clustering,” Expert Syst.With Appl., vol. 39, no. 5, pp. 5946–5959, 2012.

[75] A. C. Rencher, “Interpretation of canonical discriminant functions, canon-ical variates, and principal components,” Amer. Statist., vol. 46, no. 3,pp. 217–225, 1992.

ZHOU et al.: PROSPECT-THEORETIC MODELING OF CUSTOMER AFFECTIVE-COGNITIVE DECISIONS 483

Feng Zhou received the Bachelor’s degree in in-formation science from Ningbo University, Ningbo,China, in 2005, the Master’s degree in computerengineering from Zhejiang University, Hangzhou,China, in 2007, and the Ph.D. degree in humanfactors engineering and human–computer interactionfrom Nanyang Technological University, Singapore,in 2012. He is currently working toward the Ph.D.degree in engineering design with the Georgia Insti-tute of Technology, Atlanta, GA, USA.

His main research interests include engineeringdesign, product ecosystem design, and human–computer interaction.

Yangjian Ji received the B.S. degree in mechanicalengineering from Yangzhou University, Yangzhou,Jiangsu, China, in 1996, the M.S. degree in prod-uct design methodology from Jiangsu University,Zhenjiang, Jiangsu, in 1999, and the Ph.D. degreein mechanical engineering from Zhejiang University,Hangzhou, China, in 2004.

He is currently an Associate Professor with the De-partment of Mechanical Engineering, Zhejiang Uni-versity. He spent one year as a Visiting Scholar withthe G.W. Woodruff School of Mechanical Engineer-

ing, Georgia Institute of Technology, Atlanta, GA, USA. His main researchinterests include product system design, mass customization, and intelligentdecision making.

Roger J. Jiao received the Bachelor’s degree inmechanical engineering from the Tianjin Universityof Science and Technology, Tianjin, China, and theMaster’s degree in manufacturing engineering fromTianjin University, Tianjin. He also received the Ph.D.degree in industrial engineering from the Hong KongUniversity of Science and Technology, Hong Kong,in 1998.

He is currently an Associate Professor of Designand Manufacturing Systems with the G.W. WoodruffSchool of Mechanical Engineering, Georgia Institute

of Technology, Atlanta, GA, USA. Prior to joining Georgia Tech, he has workedas an Assistant Professor and then an Associate Professor with the School ofMechanical and Aerospace Engineering, Nanyang Technological University,Singapore. His research involves engineering design, manufacturing systems,industrial engineering, human factors, and systems engineering.