adfa, p. 1, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A context-aware customer experience management development
framework based on ontologies and computational intelligence
Hafedh Mili, Imen Benzarti, Marie-Jean Meurs, Abdellatif Obaid, Javier Gonzales-Huerta, Narjes
Haj-Salem, Anis Boubaker
LATECE Laboratory, Université du Québec à Montréal
Abstract. Customer experience management (CEM) denotes a set of practices, processes, and tools, that aim
at personalizing a customer's interactions with a company around the customer's needs and desires (Weijters
et al., 2007). E-business specialists have long realized the potential of ubiquitous computing to develop con-
text-aware CEM applications (CA-CEM), and have been imagining CA-CEM scenarios that exploit a rich
combination of sensor data, customer profile data, and historical data about the customer interactions with
his environment. However, to realize this potential, e-commerce tool vendors need to figure out which soft-
ware functionalities to incorporate into their products that their customers (e.g. retailers) could use/configure
to build CA-CEM solutions. We propose to provide such functionalities in the form of an application
framework within which CA-CEM functionalities can be specified, designed, and implemented. Our
framework relies on, 1) a cognitive modeling of the purchasing process, identifying the potential touch-
points between the seller and the buyer, and the relevant influence factors, 2) an ontology to represent rele-
vant information about consumer categories, property types, products, and promotional material, 3)
computational intelligence techniques to compute consumer- or category-specific property values, and
4) approximate reasoning algorithms to implement some of the CEM functionalities. In this paper, we
present the principles underlying our framework and outline steps for using the framework for particular pur-
chase scenarios. We conclude by discussing directions for future research.
Customer experience management (CEM) denotes a set of practices, processes, and tools, that aim at
personalizing a customer’s interactions with a company around the customer's needs and desires
(Weijters et al., 2007). Ubiquitous computing is a computing paradigm where a computation of inter-
est to a stakeholder is performed, collaboratively, by a variety of devices, often specialized, with lim-
ited capacities, that interact 'spontaneously', on behalf of the stakeholder (Palmer, 2010; Verhoef et
al., 2009). E-business specialists have long realized the potential of ubiquitous computing to develop
context-aware CEM applications (CA-CEM). Many e-business visionaries fantasize about what our
shopping experience could be like at our favorite food, entertainment, or clothing store, if the "sys-
tem" knew about our lifestyle, our age, our heartbeat, what we ate this week, the weather outside, the
people we hang out with, in real life and in social networks, the places we visit, on foot or by mouse,
what we like and what we wrote in our blogs, etc... and if all the 'things' within our environment were
connected (Weijters et al., 2007), including our cell phone, the store we walked into, the steak slab we
put in our shopping cart, and if we shared the same identity across devices, media, and applications.
This raises a multitude of questions, both for the e-commerce software vendor who wants to offer
context-aware customer experience management functionalities, and for customers of such applica-
tions (e.g. retailers), who wish to take advantage of such functionalities, including:
What customer experience management (CEM) functionality do we want to provide? Different
actions can be taken, depending on the phase of the consumer experience. For example, during
pre-purchase, we can try to anticipate a customer need ('your car is due for an oil change'), or re-
spond to a product search on the company's website by promoting a particular product. During pur-
chase, we can develop functionalities for cross-selling ('would you like a tie with your shirt?') or
up-selling ('did you check [the better] brand X?'). After the purchase, we can conduct customer sat-
isfaction surveys, or monitor customer posts on product forums, blogs, or social networks. etc.
Which information about the customer, the product, the promotional material, and the environment
do we need, in order to support the various CEM functionalities? For example, to be able to antici-
pate a customer need, what kind of information do we need to know about the customer? this de-
pends on the need. For some basic needs (e.g. food), I don't need much. For lifestyle purchases, a
lot more is needed. The same can be said about products. For some commodities, price may be all
that consumers care about. However, health- or socially-conscious consumers may care about the
production process of what they are buying, including environmental considerations, fairness, sus-
tainability, labor practices, etc
Having identified the kind of data needed about the customer to support my CEM functionalities, I
have to figure out how to capture it. This, too, involves a number of issues. Consumers are--
understandably--loathe to supply personal/demographic information, and marketers have to be
creative about weaning that information from actual or potential customers. Then, there is the
whole issue of subjective information about the consumers, such as their beliefs, values, and senti-
ments about products, processes, and issues. Finally, there are legal and ethical issues related to the
capture and exploitation of such data.
We propose to address these challenges within the context of a context aware, customer experience
management (CA-CEM) development framework that enables us to:
1) specify the CA-CEM functionalities that we wish to support within the context of a purchasing
2) translate these CA-CEM functional requirements into software specifications, in terms of required
data structures and algorithms to support the CA-CEM functionalities
3) translate those software specifications into actual code using predefined software artefacts (librar-
ies, templates, generators, etc).
Traditional so-called application frameworks embody an architecture/a design, with predefined varia-
tion points that can be instantiated for the application at hand. As such, they support only the third
step of our framework. By comparison, the development framework that we propose would cover all
the steps from business/user requirements to code.
To be able to support the requirements phase, we need to understand/delimit the problem space, i.e.
the requirements space for customer experience management. To this end, we relied on a cognitive
modeling of the purchasing process to identify the various decision points, and the decision crite-
ria/influence factors relevant to each decision point (section 3). This modeling helps us identify the
touchpoints between the seller and the seller and the customer that are needed to manage the customer
experience. In other words, this cognitive model enables us to script the interactions between the
seller and the customer in a way that allows us to read or change the mind of the customer at pivotal
moments of the purchasing experience (section 4).
The cognitive/functional design of the interactions between the seller and the customer relies on:
1) the identification of the relevant information about, a) the customer (customer profile), b) the
products, and c) communication content between the customer and the seller-- e.g., promo-
2) the selection or design of algorithms and tools to fill that information from the available
3) the selection and design of algorithms to customize the interactions between the seller and the
customer based on the information.
The relevant information is presented in the form of ontologies (section 5) to be specialized or
instantiated for specific purchasing scenarios (section 7.3). We rely on computational intelligence
techniques to, 1) fill out some of the property values, including subjective information from, and
about the customer (e.g. preferences, sentiments, or beliefs, see section 6.1), and 2) customize the
interactions between seller and customer through product recommendation, targeted marketing, and
the like (Section 6.2).
This paper presents the principles behind our approach. Section 2 introduces the process view of
CEM, using a retail example. Section 3 describes the purchasing process, from an operational and a
cognitive point of view. It identifies the different steps of the purchasing process, and the various fac-
tors that can influence customers in their decision making. Section 4 lays the foundations of our
framework by going from the cognitive model of the purchasing process to, a) a generic customer
experience management pattern, and b) purchasing processing step-specific specializations. These
patterns helped design the ontologies needed to implement them, which are presented in section 5.
Section 6 shows how computational intelligence techniques can be used to fill out consumer profiles,
and to customize the interactions with them. Section 7 shows how this framework can be instantiated
to handle a particular purchasing scenario. We conclude in section 8 with a discussion of the work
remaining to be done.
2 A process view of customer experience management
In this section, we use an example of a purchasing scenario that illustrates some of the possibilities
and requirements of a CA-CEM framework. We start by presenting the scenario, and then present our
process view of customer experience management.
2.1 A CA-CEM Scenario
Chris is a thirty something young urban professional who walks into her favourite grocery store where
she usually shops. As she drops items in her shopping cart, the food labels are automatically displayed
on her (latest) iPhone. As she drops a box of crackers, she gets a warning, because of its high-level of
sodium--given her family history of blood pressure. She walks through the produce section, and gets
notices about latest arrivals of fair trade certified products. She is pleased, of course, being an active
member of Equiterre1, but wonders if every shopper gets notified. Walking into the meat section, her
attention is drawn to a special on lamb chops, that her significant other enjoys immensely. She picks a
rack and drops it into the shopping cart. She gets wine suggestions: two thumbs up for a Syrah from
northern Rhône, and one thumb up for a Shiraz2. The Shiraz wins with the ongoing special on Austra-
lian wines. Cheese with that? A French baguette? In the seafood section, Chris picks up a slab of Tuna
steak. She gets a warning to the effect of having consumed big fish already four times already this
week3. While getting toothpaste, she gets a notification to the effect that size 4 diapers are on special.
Considering that she has been buying size 3 diapers for the past six months, it is about time she
switched to size 4!
Some might find this 'idyllic' shopping experience frightening, and rightfully so. We do not presume
that this scenario is desirable: we will simply explore the capabilities that make it possible. In particu-
lar, we explore the kind of data/knowledge needed about the products and the customer to make such
a scenario possible. We will do this step by step:
1. ".. As she drops items in her shopping cart, the food labels are automatically displayed on her
iPhone". This, of course, is easy enough, and can be done in many different ways, either by having
short range RFIDs, which are probed by emitters in the cart, or simple bar codes scanned by de-
vices upon being dropped in the cart. The technology for this exists today in many different forms.
2. " As she drops a box of crackers, she gets a warning, because of its high-level of sodium--given her
family history of blood pressure". This requires 'the system' to have some sort of a medical profile
of the customer: what conditions they already have, or they are predisposed to.
3. "She walks through the produce section, and gets notices about latest arrivals of fair trade certified
products". This notification depends on the system having access to the customer's social values.
These values could have been entered directly by the customer when signing up for a particular
service or social network. Alternatively, we can infer the customer's interests, based on their likes
(Facebook profile), on the groups they are following (LinkedIn), on their membership to various
advocacy groups that promote specific values (e.g. Equiterre), or on posts in various social medias4.
4. "Walking into the meat section, her attention is drawn to a special on lamb chops, that her signifi-
cant other enjoys immensely". Knowing that someone likes a particular food product is easy
enough. However, underlying this recommendation is a deeper understanding of human relation-
ships, and what they entail: 1) people live with their spouses, and 2) as such, they eat together. If it
is Chris's office buddy, or sibling who likes lamb chops, the system should make no such recom-
5. "She gets wine suggestions: two thumbs up for a Syrah from northern Rhône, and one thumb up for
a Shiraz. The Shiraz wins with the ongoing special on Australian wines. Cheese with that? A
1 www.equiterre.org, whose mission statement includes "Equiterre helps build a social movement by encouraging individu-
als, organizations and governments to make ecological and equitable choices, in a spirit of solidarity. We see the everyday
choices we all make - food, transportation, housing, gardening, shopping - as an opportunity to change the world, one step at
2 Gracieuseté of an entry in the INTOWINE web site, http://www.intowine.com/best-wine-pair-lamb-chops, accessed
3 Big fish, because they are higher up in the food chain, contain more toxic substances such as mercury.
4 More difficult, with big data techniques such as sentiment analysis and the like.
French baguette?". This is standard market basket analysis, combined with recommender function-
ality: it does not rely on any data specific to the customer.
6. "Chris picks up a slab of Tuna steak. She gets a warning to the effect of having consumed big fish
already four times already this week". Such a functionality depends on the system embedding a
number of health advisories, and a history of the customer's purchases.
7. " she gets a notification to the effect that size 4 diapers are on special. Considering that she has
been buying size 3 diapers for the past six months, it is about time she switched to size 4". This is
an extreme/fine-grained case of family life cycle marketing. The concept of family lifecycle market-
ing recognizes that families go through a predictable set of stages (family lifecycle) during which
their needs, means, and decision patterns, evolve (Wells et al., 1966). Marketers take advantage of
this lifecycle to, 1) better identify the target market segment, and 2) adapt their marketing message
accordingly. In our scenario, not only does our 'system' recognize that Chris is in the so-called 'full
nest I' stage5-- which can be inferred from her track record of repeated diaper purchases-- but it is
also predicting her child's progression through her/his lifecycle stages
To summarize, to make this idyllic (or frightening) purchasing scenario possible, the system needs:
electronically accessible detailed product information, including nutritional information (point 1.),
production mode, toxicity (point 6.), and value-based assessments and certifications (point 3.)
product associations (point 5.)
a detailed customer profile, including medical history (point 2.), demographic data and lifecycle
stage (point 7.), relationships (point 4.), tastes (point 4., re-spouse), beliefs and values (point 3.),
a history of purchases (point 6. and 7.).
Note the history of purchases can be used us to infer-- or more accurately, take a guess at-- other cus-
tomer profile data, such as tastes or beliefs and values. For the time being, we will not worry about
how the different pieces of information can be obtained. That concern will be addressed later.
In the next section, we analyze these different pieces of information within the context of a process-
oriented view of the interaction between an organisation offering a product, and its customers.
2.2 A process view of CEM
Customer Experience Management aims at managing the interactions between a company and its cus-
tomers. An actual or potential customer interacts with a company to fulfill a particular need, in the
form of a product or service provided by the company. A company interacts with its customers, actual
or potential, because that is its raison d'être: fulfilling the needs of its customers. A for-profit organi-
zation gets paid in return for the fulfillment of that need at a price that is higher than its production
cost. A public service organization (e.g. a government department) draws its value from the fulfilment
of the needs of the citizenry.
5 Identified as "young married couple with dependent children".
A customer has a 'life of his/her own' pursuing his/her objectives of survival and pursuit of happiness.
In the process of pursuing those objectives, they have needs that can or need to be fulfilled by a com-
pany. Customer experience deals with the interactions between the customer and the company around
the act of need fulfillment: the product or service sale. Fig. 1 illustrates this interaction. The family
lifecycle theory recognizes that the needs of people evolve during their life, as they enter different
stages of the cycle. Those needs depend on many factors, including the processes underlying those
stages (e.g. raising children), as well as the means that are typically available to customers at those
Fig. 1. Customer experience as the set of interactions between a company and a customer around the fulfilment of a cus-
In the next section, we will look at the mechanics of the interactions covered by the customer experi-
ence by identifying the different stages of a purchasing process, and the various factors that influence
the purchasing decision.
3 Understanding the purchasing process: a cognitive approach
Consumer behavior has been studied thoroughly by marketers and social psychologists trying to un-
derstand its mechanics. They use the term ‘consumer’, in the broad sense, where the object of con-
sumption can be a product/service (food, jeans, a cell phone service package, a car, a house), a behav-
ior (exercising, dieting), or a belief (social values, political affiliation). At a fairly basic level, con-
sumption is a conscious purposeful behavior, whose goal is to address a need or a desire. A number of
psychological models have been proposed, including the theory of reasoned action (Fishbein et al.,
1975), the theory of planned behavior (Ajzen, 1985), the MODE model (Fazio, 1986), the theory of
trying (Bagozzi et al., 1990), the theory of self-regulation (Bagozzi, 1992), and subsequent variations
thereof. The theory of reasoned action (TRA) perceives all actions as purposeful behavior that starts
by forming an intention to act, followed by the performance of the action itself (Fishbein et al., 1975).
TRA recognizes that intentions are influenced by two factors, namely, the attitude(s) towards the ac-
tion, and the so-called subjective norm; the attitude towards the action is defined as the perceived like-
lihood of some outcome occurring. The subjective norm refers to the actor/consumer's perceived be-
lief that (certain) people expect him/her to behave/act or not act. To use a simple example, I intend to
of happiness, etc)
interactions related to the ful-
fillment of the customer needs
buy Nike shoes, because, a) I think I will look cool (expected outcome for me), and b) I think that my
peers expect me to wear sneakers (‘peer pressure’). Researchers and experiments have poked holes in
this theory, which was amended to include other salient influences, such as an actor’s belief in their
ability to complete the action (perceived behavioral control), leading to the theory of planned behav-
ior (Ajzen, 1985), and to account for impulsive action/consumption, leading to the MODE (motiva-
tion and opportunity as determinants …) (Fazio, 1986). Other amendments took into account the
complexity of the actions needed to consume, where the actions towards the consumption goal, be-
come, themselves, goals of their own, leading to the theory of trying (Bagozzi et al., 1990). And so on
and so forth.
For the purposes of this paper, we will be content to present the synthesis presented by Bagozzi et al.
(Bagozzi et al., 2007), which integrates all of the influences that have been identified by researchers.
The model is shown in Fig. 2. We will briefly explain the various steps, and the influencing factors;
the significance of this model for customer experience management will be explored in section 4.
intention Trying Goal-directed
Perceived behavioural control
and motivation to
moral values and
Fig. 2. A comprehensive model of consumer behavior adapted from (Bagozzi, 2006)
First, we start with general observations/principles. First, there are different paths through this proc-
ess, depending on the type and complexity of consumption decisions. For example, habitual or low-
risk consumption activities (e.g. grabbing a carton of milk) involve deliberation or planning. Second,
social psychologists make the distinction between desirable goals and desirable behaviors, with the
former often preceding the latter. For example, my goal could be to lead a healthier life, which repre-
sents a desirable end state. This goal may entail one or many alternate (set of) behavior(s), such as
dieting and exercising, where the goal is a behavior6. For the sake of simplicity, we represent them as
two independent "process paths", ignoring the precedence relationship that can exist between goals
and behaviors. Third, this model makes a distinction between desires, intentions, plans, and actual
a consumer may have several desires (goal desire or behavior desire in Fig. 2), but intends to pursue
only one (goal intention or behavior intention in Fig. 2). I would love to have a car to take my kids to
school, a motorcycle to take leisurely rides during the week-end (two goal desires), but I am going to
stick with buying a car (goal intention). The same applies to behaviors--otherwise known as new year
having decided to pursue a particular goal (goal intention) or behavior (behavior intention), the con-
sumer needs to figure out how to achieve that goal, i.e. needs a plan of action to achieve the goal. For
example, having decided to purchase a car, I need to figure out how to do it. This is as implementa-
tion intention in Fig. 2
having devised a plan, the consumer then executes the plan, by performing the actions of the plan.
This is covered by the two steps, trying and goal-directed behavior in Fig. 2.
We will comment on some of the finer distinctions below.
In the following, we will describe each step of the process of Fig. 2, and for each such step, describe
the relevant influence factors:
1) Goal desire: corresponds to identifying the various needs and desires of the customer. Several
factors influence the 'setting' of a goal:
a) Goal feasibility: I may fantasize about owning an executive jet, but I won't give it serious
consideration/consider it as a potential goal because it is unattainable.
b) Anticipated positive emotions: this represents the perceived reward of attaining a particular
goal, and it is a combination of, i) positive emotions resulting from success ('how good I will
feel'), and ii) the expectation of success (how likely I am to succeed)
c) Anticipated negative emotions: this represents the perceived penalty of failing to attain the
goal, which is a combination of, i) negative emotions resulting from failing (how bad I will
feel if I fail), and ii) the expectation of failure (how likely I am to fail).
d) Outcome expectancies: the expected outcome of the pursuit of the goal (success vs. failure)
e) Social identity: an individual's membership to a particular group, and the "emotional and
evaluative [positive or negative] significance of this membership" (Bagozzi et al., 2007b).
6 A good number of the studies that helped build these models concern behaviors like smoking, drinking, dieting, recycling,
exercising, etc. Thus, the 'consumption process' under study actually starts with the behavior desire.
f) Frequency of past behavior: this was not included in Bagozzi's framework7, but is an impor-
tant aspect of the theory of trying, which is one of the foundations of this framework. This ba-
sic idea is that customers may undergo the complex cognitive processes involved in goal set-
ting once, or the first few times, but after that, they make a cognitive shortcut: "I have thought
this through many time, and found it worth it to consider or pursue this goal"
2) Goal intention: Goal desire is concerned with the desirability of goals, but not with the decision
to pursue them: this is step is where the decision is made.
3) Behavioral desire: Researchers have identified six influences : 1) social identity, 2) outcome ex-
pectancies, 3) frequency of past behavior, 4) attitudes, 5) subjective norms, and 6) perceived be-
havioral control and self-efficacy. The first three were discussed above. We discuss the remaining
a) attitudes: these were identified by the theory of reasoned action (TRA, (Fishbein et al., 1975))
as influencers of the pursuit, or not, of some actions. A positive attitude about an action leads
the consumer to act, whereas a negative attitude inhibits him/her. The attitude towards an ac-
tion is a combination of, i) my belief about what would result from the action (what is the
outcome), and ii) my evaluation of that result. For example, if I buy Jeans, I believe that I will
look fashionable (the outcome), and, I like being fashionable.
b) subjective norms: were defined by Ajzen & Fishbein as an individual's perception of the so-
cial pressure to perform, or not perform, an action (Ajzen et al., 1980). That pressure is a
combination of, i) how confident I am that my spouse/colleague/peer group expects me to
perform the action (a probability), and ii) the perceived reward (or penalty) from complying
(or not) with that expectation. For example, I am 90% confident that my colleagues expect
me to wear Jeans, and, if I don't comply, I believe that I will be eating alone.
c) perceived behavioral control and self-efficacy: perceived behavioral control was introduced
by the theory of planned behavior as a determinant in the decision to undertake or not, some
actions/behaviors. It is defined as an individual's perception of how easy or difficult it is to
perform a particular behavior (Ajzen et al., 1986). It can be thought of as combination of, i)
my belief in me possessing a factor I think is needed to perform the behavior, and ii) my be-
lief in how important that factor. For example, to stop smoking, I think that will power is eve-
rything, and I conceded that I don't have strong will power. For the purposes of this paper, we
will take perceived behavioral control and self-efficacy to be synonymous8.
4) implementation intention: this is the planning phase. Once I have chosen a goal or a goal-directed
behavior to perform, I plan for it. This is particularly relevant when the purpose of my consump-
tion is an end state, i.e. when we are dealing with a goal desire as opposed to a behavior desire. In
such a case, implementation intention consists of identifying the steps/individual actions I need to
perform to reach my goal-- which can themselves become goals, in their own right, if their per-
7 Fig 4.5, page 97, in (Bagozzi et al, 2007b)
8 there is some debate as to whether perceived behavioral control and self-efficacy are the same thing. Ajzen thinks so (e.g.
http://people.umass.edu/aizen/faqtxt.html). Armitage & Conner think not (Armitage et al., 1999).
formance is problematic (Bagozzi et al., 2007b). The planning phase is influenced by the fre-
quency of past behavior: if I solved the problem (of goal attainment) several times before, I can
reuse the same solution (implementation plan). We discuss the second influencing factor below:
Second-order moral values and self-evaluative standards (Bagozzi et al, 2007b). Values are defined
as the criteria or frame of reference by which people justify actions and 'judge' actions and people
(Schwartz, 1992). Self-evaluative standards represent how consumers see themselves, or what they
want to become (Bagozzi et al., 2007b). Our shopper from section 2.1 is a socially responsible con-
sumer, and that is why the 'grocery store' pitched the latest arrival of politically correct products. Fig.
2 shows the moral values and self-evaluative standards are influenced by the consumer's social iden-
5) Trying: within the context of the original theory of trying, the act of trying was thought of as "a
singular subjective state summarizing the extent to which a person believes they have tried or will
'try' to act"(Bagozzi et al., 2006). The definition was later extended to the actual execution of the
plan established in the previous step, where each action in the plan can be problematic, and be
forestalled (Bagozzi et al., 1990). "Trying" involves monitoring progress towards the objective,
and making adjustments to the plan, as appropriate. Bagozzi uses the example of showing appre-
ciation to a friend at their birthday (a goal intention) by buying her/him a gift (a behavior desire),
which involves of a plan of going to a shopping mall, selecting stores to visit, and browsing
through merchandise to select a gift within the selected budget (implementation intention), to ac-
tually executing the plan (trying), which involves going to the mall, etc. (Bagozzi et al.,
2007).The consumer is not guaranteed to be able to do any of the steps required to make the pur-
chase, including: 1) getting to the mall (transportation problem), 2) finding a set of stores of inter-
est (not the right kind of merchandise), and 3) finding merchandise that fits the taste of the gift re-
ceiver, and the budget of the gift giver. The trying phase is influenced by the frequency of past
behavior, discussed above, and the recency of past behavior, discussed below.
a) Recency of past behavior: within the context of the theory of trying (Bagozzi et al., 1990), the
recency of past behavior can influence current behavior in the same way that the frequency of
past behavior does, because the behavior is still 'fresh' in the consumer's mind
6) Goal-directed behavior: refers to the 'final act' in the consumption process, for example making a
7) Goal attainment/failure: this is the step where the consumer assesses the extent to which they
have reached their goal.
8) Feedback: based on the previous assessment, the consumer can adjust any of the choices or ac-
tions made in the 'consumption process', including the choice of goals to pursue.
This 'comprehensive model' of consumer behavior, proposed by Bagozzi (Bagozzi et al., 2006) with
very minor adaptations of our own, is more or less the result of 'merging' consumer behavior models
pertaining to different kinds of consumptions, for which some of the steps, or the influencing factors,
may be irrelevant. Our framework for CA-CEM should provide guidance (or tools) for customizing
this process for specific consumption or CEM scenarios; that is discussed in section 7.1.
4 A cognitive approach to context aware customer experience management design
4.1 A customer experience management pattern
Section 2.2 showed a view of CEM as managing the interactions between two processes executing on
behalf of two entities, the enterprise/seller, and the customer, each with their own objectives. Those
interactions centered on the act of consuming. Section 3 presented the different steps of the consump-
tion process and identified the factors that influence each step, without worrying about interaction
points between the enterprise and the consumer. In this section, we put it all together.
Fig. 3 shows the idea. Consumers are 'active systems' (living organisms) whose internal processes (to
stay alive, pursue happiness, etc.) require a number of resources (food, clothing, transportation means)
or conditions (fulfillment, happiness, etc.), triggering consumption processes to replenish the re-
sources ("we are out of cereal") or to attain those conditions ("I need to go/dine out"). Those con-
sumption processes involve a number of steps.
Fig. 3. Towards a systematic view of consumer experience management
The decisions and choices taken in those steps are influenced by a number of factors, which have been
identified by psycho-sociological studies of the consumer, and discussed in section 3. Some of the
steps of the consumption process involve interactions between the enterprise and the consumer. The
interaction can be triggered by the consumer, for example by accessing the enterprise's web site, or
triggered by the enterprise, by communicating with the customer, nominally, or through advertising.
To the extent that some of the steps of the consumption process involve choices and decisions that are
influenced by a number of psycho-sociological factors, the enterprise gains from, minimally, knowing
the 'values' of those factors. Some of these influence factors are malleable, such as the anticipated
positive emotions, and the enterprise may gain from modifying their values to its advantage, for ex-
ample, by showing potential consumers how good life would be, if they purchased its products, or
how easy it is to get approved for credit to make the purchase (e.g. perceived behavior control). Such
factors are referred to as read-write influence factors in Fig. 3, as opposed to read-only (RO) influence
factors, that enterprises cannot modify or act upon. Further, while all the read-write influence factors
are potentially actionable from a CEM point of view, some/many may not be worth pursuing.
4.2 Towards an integrated view of customer experiences
Fig. 4 shows a first cut naive application of the pattern shown in Figure 3 to all the steps of the con-
sumption process, incorporating the full roster of influence factors identified by the literature on social
psychology of consumer behavior (Bagozzi et al., 2006). However, such a view is reductive in the
1) a seller, acting through its e-commerce system, does not necessarily accompany a (potential)
buyer through all the steps of the purchasing process. This is particularly true for the early steps of
the process, such as the "goal desire" stage. This is not to say that sellers have no influence on
"goal desire": quite the contrary. For example, the tobacco industry, the beer industry, and the
fashion industry, to name a few, all act upon the anticipated (positive) emotions-- and to a lesser
extent, social identity-- associated with the consumption of their products; it is just that this influ-
ence does not happen within the context of a one-to-one interaction between the customer and the
seller's e-commerce system
2) as a corollary of the first observation, different channels will typically accompany different sub-
sets of the purchasing process-- often corresponding to different entry points into the purchasing
process. For example, I can get into an e-commerce site through a web search for a product,
where the returned page will display the product specs, and propose to add it to a shopping cart,
thereby skipping "goal/behavior desire", "goal intention", "implementation intention", and part of
"trying". However, a walk into a particular store will instantiate/trigger the full cycle, starting with
a welcoming display featuring (promoting the desirability of) a specific product.
3) a second corollary is related to the fact that a consumer may interact with several sellers for some
steps of the process. This may be the case early in the process before the customer commits to a
product and a seller. For example, you walk into a shopping mall and the location-based services
on your phone send you advertisements from the different restaurants or shoe stores in the mall.
Knowing that you are in competition with other contenders will impact the way you interact with
the (potential) customer.
4) a lot happens under the steps 'Trying' (executing a 'consumption plan') and 'Goal directed behav-
ior', that may include information search, comparison shopping, consulting product reviews, and
so forth i.e. lots of sub-steps that do involve interactions between seller and consumer. The pattern
of Fig. 3 will need to be applied to those sub-steps.
With this is mind, in the next section we do study the various steps of the purchasing process and see
whether and how we can apply the pattern of Fig. 3.
Fig. 4. A naive view of customer experience management
4.3 Customer experience management as service design
The design of E-commerce applications, in general, and customer experience management applica-
tions in particular, can be studied within the context of service design. We all have an intuitive un-
derstanding of what services are (as opposed to goods), and they include things such health care,
education, transportation, hospitality, phone or telecommunications, etc. Wikipedia defines a ser-
vice as an intangible commodity, where a commodity is a "marketable item produced to satisfy
wants and needs"9. Service delivery typically involves a process that orchestrates people, compo-
nents, structures, and automated systems--typically IT-- to satisfy the needs/wants of the customer.
Service design is concerned with the selection of the people, components, structures and automated
systems needed to deliver the service and the design of the interactions between them in a way that
is economically viable for the service provider and that maximizes customer satisfaction. Service
design theory involves a number of disciplines, including marketing, operations management, organ-
izational design, and technology (see e.g. (Cook et al., 2002; Hill et al., 2002; Ostrom et al., 2010)).
The outputs of service design (good and bad) include the work organization at your local bank
branch, your Department of Motors and Vehicles bureau, or your favorite department store.
9 See https://en.wikipedia.org/wiki/Commodity, accessed on 18/8/2015.
Execution and control)
Moral values &
Frequency of past
recency of past
There has been a growing interest in the research community to look at experience management
within the context of service design (see e.g. (Cook et al., 2002; Teixeira et al., 2012)), with the intent
of incorporating customer experience requirements along with the other requirements that drive the
service design. Teixeira et al. introduced customer experience modeling (CEM) as a "model-based
method .. to represent and systematize customer experiences for service design."(Teixeira et al.,
2012). CEM draws on Constantine's Human Activity Modeling (HAM) (Constantine, 2009), on cus-
tomer experience requirements, themselves from Mylopoulos et al.'s goal-oriented requirements
analysis (Mylopoulos, Chung, & Yu, 1999), and Patricio et al.'s multi-level service design (Patricio,
Fisk, Falcao e Cunha, & Constantine, 2011). Figure 5 shows the concepts of CEM and their relation-
ships. Note that customer experience requirements apply to both activities and the context within
which these activities take place, including other actors (e.g. a sales clerk), artifacts (e.g. a passive
display), and 'technology enabled systems', including computing devices, sensors, and the like.
Figure 5. Incorporating customer experience requirements in service design. From (Teixeira et al., 2012).
An important aspect of service design is customer scripting (see e.g. (Cook et al., 2002)), which is the
specification/design of the interaction process, between a customer and a service provider, needed
to deliver the service. This involves the careful design of both, a) the individual interactions and in-
teraction touchpoints, and b) the orchestration of such interactions. As an example of a), we explore
different ways for the consumer to identify themselves, and select the easiest ID method. An exam-
ple of b) is whether to request identification from the beginning of a service script (which would re-
sult into more appropriate/customized service interactions), or adopt a light process, and wait until
the customer proceeds to checkout, to request identification. The customer experience modeling
method is, for the most part, descriptive, providing modeling concepts and notations. However, it
offers little guidance as to how to design a service in a way that it conforms to the customer experi-
ence requirements. Cook et al. identified some of the human and subjective issues surrounding cus-
tomers' appreciation of their service encounters/experiences, independently of the intrinsic qualities
of the product/service rendered, and of the customer specifics. For example, one of the script design
guidelines, based on psychological studies, suggests that we design a customer script that puts un-
pleasant interactions/steps first, and ends with a high note (Cook et al., 2002).
i s sur r ounded by
Within the context of service design, the description of the purchasing process shown in Fig. 2. pro-
vides a high-level solution space for consumer scripting. As mentioned in section 3, our e-commerce
(or experience management) application will have to implement a sub-script (sub-process) of the
process described in Fig. 2, depending on the type of purchase (carton of milk, versus car) and on
the channel (brick or click). The pattern of Fig. 3 shows us what to look for to design the individual
interactions/steps of the script. We see next how to turn the psychological influences described in
Fig. 2. into specific actions/interactions to incorporate into a consumer script.
4.4 Turning psychological influences into actionable consumer experiences
In this section, we attempt the use of the pattern of Fig. 3 to the steps of the purchasing process de-
picted in Fig. 2. We start by outlining the principles (section 4.4.1). We then look at specific exam-
ples (section 4.4.2).
We start by categorizing the purchasing process steps, and their influences, and then talk about the
different ways we can exploit those influences.
Roughly speaking, the purchasing process involves two kinds of activities, 1) internal cognitive activi-
ties, taking place within the consumer's head, and 2) external, possibly physical, activities, some of
which involving explicit interactions between the customer and the seller/service provider. Examples
of cognitive activities include goal/behavioral desire, goal intention (selecting one worthwhile goal
among many) and goal implementation (planning a way to achieve the goal). Examples of external
activities include trying (plan execution) and goal directed behavior. An example of external activity
that involves an interaction with the seller is doing an on-line product search on the seller's portal,
asking an in-store sales clerk for advice, or checkout (online or in store).
With regard to the influence factors, we categorize them based on two dimensions:
1) subject/scope: customer-specific factors (socio-demographic data, social identity, moral values
and personal standards), product/service-specific factors (product specifications), and factors re-
lated to the <customer, product> relationship (anticipated positive emotions, anticipated nega-
tive emotions, history of past behavior)
2) type: objective/factual factors (customer socio-demographic data, history of past behavior,
product specifications) versus subjective/emotional/perception-based factors (attitudes, subjec-
tive norms, perceived behavioral control).
From a service script/interaction design point of view, for each activity of the purchasing process, we
can do different things, depending on the kind of activity, and the kind of factors influencing that
1) Purely cognitive activities. These activities happen inside the consumer's head (e.g., goal inten-
tion), and do not require interaction with the seller/service provider. Thus, we need to provoke
an interaction, as part of our service script design. This interaction can take place:
a) prior to the activity: to modify the subjective/emotional/perception-based factors. For ex-
ample, prior to goal desire/goal intention, I can strengthen the anticipated positive (nega-
tive) emotions associated with a goal, to make the customer desire it and select it over other
b) after the activity: in this case, we cannot influence the activity, but we can at least hope to
get a reading on the choices made so that we can better prepare the response of the service
provider to the subsequent activities. For example, as a retailer, I may not be able to influ-
ence 'goal intention' (selecting one goal to pursue, among many desired, e.g. which item of
clothing you came into the store to purchase, among all the ones we offer). But if I can tell
(or ask!) which item you came in for, I can direct you.
2) External activities that normally involve an interaction with the seller/service provider. In this
case, the activity is part of the normal script, and the customization concerns the information
content, and how it is delivered, which are adapted to what we know about the customer. An
example of such activities includes searching the seller's online catalog. If we know who the cus-
tomer is, we can refine the search query, and customize the presentation of the results
3) External activities that do not normally involve an interaction with the service provider. Here
too, we need to provoke an interaction with the customer, either before the activity takes place,
to influence its outcome, or, failing that, after it has taken place, to get a reading on what the
customer did, to plan the subsequent steps of the process/service script. An example of such an
activity is searching for product reviews (pre-purchase), or entering a product review (post-
purchase), in a third party site.
The difficulty with the activities that do not (normally) involve an interaction with the seller/service
provider (via their e-commerce software) is to guess where the customer is, along the purchasing
process. A seller may adopt a default strategy of assuming that you are at the (pre) goal desire stage,
and send advertisements your way, until/unless you do something specific. However, when we get
into the intermediate stages (e.g 'goal intention', 'goal implementation', and the various sub-
steps/sub-stages of 'trying'), things get tricky. Not knowing where the customer is in the purchasing
process can be challenging even when the customer initiates an interaction. When you walk into a
dealership, the salesperson will first ask you whether you are looking for a model in particular, or
just looking, and if you answer the latter, they may go back to playing Solitaire. Similarly, when you
do a product search on an online catalog, you may be looking for a particular model whose model
number you forgot, or looking for any product that matches your specifications; depending on the
situation, there are different strategies for completing a sale!
In this section, we show how to use the concepts and principles described above (section 4.4.1) to
design interactions for specific purchasing process steps. The purpose here is to illustrate the kind of
analyses a 'customer experience designer' needs to make. The final 'answer' would rely/depend on
marketing knowledge. This analysis will also help us start determining how the identified influence
factors translate into data that an e-commerce application can manipulate; section Error! Reference
source not found. will propose a first-cut ontology/metamodel to support the experience manage-
ment functionalities illustrated throughout the paper. We start by a thorough analysis of the 'goal
desire' stage, and then present pointers related to other stages.
At a basic level, goal desire is triggered by the lack or deficit of physical resources ("we are out of
milk") and physical ("I am out of shape") and mental states ("I need some fun") (see Fig. 4). With
the exception of biological needs, for which we need no prompting, companies/marketers can cre-
ate a need10 or strengthen a need through advertising. According to Bagozzi's framework (see Fig.
2), the strength of such desire is influenced by anticipated emotions (positive and negative), out-
come expectancies, and social identity. This raises two kinds of questions:
1) what is the appropriate advertising message content to strengthen the desirability of a goal. In
particular, we need to think of the extent to which a (better) knowledge of the customer profile
can help customize/select the appropriate message
2) how to deliver that message within the context of a CEM application, considering the stage of
the purchasing cycle we are at.
At a basic level, we can strengthen the desirability of a product/service or behavior by projecting
images that, a) strengthen the anticipated positive emotions ("how good it will feel to own this
product or adopt this behavior"), b) strengthen the anticipated negative emotions ("how bad I will
feel by not owning the product or not adopting the behavior"), and c) show how easy it is to succeed
(in acquiring the product or adopting the behavior).
In what ways does our knowledge of the customer make such images more effective?
1) by depicting someone the customer can identify with. If we want to "sell" the latest Android
smart phone to a teenager or professional, we present an advertisement that features a teen-
ager, or professional, respectively. Idem with gender, ethnicity, geography, etc11.
2) by sending images of more specific goals, appropriate for their category/social identity. While
owning the latest and greatest in smart phones may be a broadly shared desire, the advertising
message could focus on social media and multimedia capabilities, when targeted towards teen-
agers, or office productivity tools, when targeted towards professionals. Similarly, if a product is
too expensive for my social category, the marketing message can stress the possibility or ease
with which I can finance the purchase-- playing on outcome expectancy
3) by sending images of goals we know the customer has not yet achieved. My cell phone provider,
or the dealer who sold me my last car, knows the features of the current product I own, and
knows which feature(s) to highlight or stress in the marketing message.
10 Marketers would disagree with the manipulative "create": they prefer the term "recognize".
11 Typically, marketers produce different variants of the same marketing theme, aimed at different populations
Having determined/selected the advertising material, we need to determine ways to deliver it. This
depends on a combination of two things, 1) whether the customer is known to the seller, and 2) the
channel. For example, if the seller already knows the customer, then we have several possibilities:
1) we can push the advertising material through an unsolicited e-mailing campaign (on-line)
2) we can present the material to the customer when they log in (on-line)
3) when the customer crosses the threshold of the store, and s/he is positively IDed, the message is
displayed on an appropriately located monitor (in-store)
If we don't know the customer, then we broadcast the message variants, both through the company
portal and in-stores.
In summary, for goal desire:
we can act upon the desirability of goals through advertising
knowing that anticipated positive emotions, anticipated negative emotions, and outcome expec-
tancies influence the desirability of a goal helps determine the orientation or content category of
the message carried by the advertising material
knowing the customer (social identity, history of consumption) can help customize the message,
the channel determines the advertising material delivery method.
Similar analyses can be performed on the subsequent steps of the purchasing process. We provide
some highlights for some of those steps:
Goal intention: here, the issue is of selecting a goal to pursue, among many that are deemed
desirable. This is another internal cognitive activity, that is practically indistinguishable from
'goal desire'. The priority of goals depends on means, lifecycle stage, values, social identity, etc. I
can influence goal selection through advertising by matching advertising content to consumer
profile. If I am a specialty outdoor equipment manufacturer, and I am talking to a young person
who can afford either a new car or a (top of the line) mountain bike, I can show her/him an ad
that features healthy and handsome young people taking their fancy bikes to an extreme moun-
tain trail in a cheap ride, or even hitchhiking there!
Goal implementation. This is the planning stage. Planning makes sense in a multichannel shop-
ping experience. For example, I have already made the decision to buy a new fridge, and I am
planning my shopping experience. How can a retailer influence the planning stage to his advan-
tage? Several general strategies include: 1) by offering a simpler shopping experience (e.g., order
from a catalog, or a better online ordering system), 2) by making sure that any shopping plan in-
cludes an interaction with me ("we offer the best price guaranteed", or "we will best the lowest
price you will find by 5%"). I can also take advantage of knowledge about the customer: 1) loca-
tion, by suggesting stores to visit to shop or pick up their merchandise. 2) shopping style, by acti-
vating/proposing appropriate shopping scripts following an on-line search.
Trying. This is the execution of the consumption plan (goal implementation). If the plan includes
in-store browsing, then we can think of the CA-CEM scenario presented in section 2.1, and vari-
ous product recommendation strategies ... especially if we know who the customer is! Imagine
walking into a department store and having your profile displayed on the PDAs of nearby sales-
people, who call you by name, and direct you to your style section.
Through these analyses, we do not pretend to provide answers that are correct, from a marketing
point of view. However, we hope to provide the design vocabulary that service designers (including
marketing specialists) can use, to design customer experience management-enabled services.
5 Ontologies for context-aware customer experience management
The discussion of section 4 showed how the combination of the cognitive modeling of the purchas-
ing process, and service design theory, can help us in:
1) identifying the steps (cognitive or physical) of the purchasing process that, either naturally lend
themselves to an interaction between the seller and the customer, or would benefit from provok-
ing such an interaction
2) selecting the modalities for such interactions, depend on the stage of the purchasing process,
and the purchasing channel
3) selecting the contents of such interactions, depending on our knowledge of the customer and the
In this section, we focus on the last aspect, the modeling of the knowledge we need to have about
the customer and the products.
We present the required knowledge models in the form of ontologies. We used the term ontology in
two complementary ways:
1. We use it in the computer science/information systems sense, i.e. as specification of a set of
representational primitives used to describe/model a domain of knowledge [Gruber, 2009]. If we
think in terms of UML and MOF, the above definition corresponds to the UML metamodel (M0),
i.e. the language used to define the UML modeling language--itself a subset of UML. An ontology
can also be understood in terms of description logics, where an ontology language can be seen
as the realization of an underlying description logic.
2. We also use in the sense of reflecting a shared conceptualization of a domain.
Indeed, within the context of our context-aware customer experience management (CA-CEM) de-
velopment framework, we want to specify the "modeling ingredients" that analysts can use to rep-
resent the user and product profiles, say, required by their purchase scenarios. We also use the term
ontology in the sense of representing shared knowledge about domains-- in this case, marketing
knowledge as it pertains to customer experience management. Such marketing knowledge may
include things such as known customer categories, and their likes, tastes, and spending habits,
known product categories, and their appeal, etc. However, to specify such knowledge, we need rep-
resentational primitives--hence the complementarity of the two perspectives.
In this section, we focus on the representational primitives needed to represent the data/knowledge
about customers, products, promotional material, and the like, that may be needed, within the range
of possible purchasing scenarios, and CA-CEM functionalities.
5.1 Consumer data
Knowing the consumer is key to successful consumer experience management. As sections 2 and 3
showed, there is a lot of relevant data. The different kinds of data are discussed in this section.
As mentioned in section 4.1, the influence factors that are relevant to the consumption process steps
represent information that consumers use to select their choices and make their decisions. Thus, com-
panies benefit from knowing what that information is, and should incorporate it in the consumer's
profile. Let us take a couple of examples from the scenario of section 2.1:
1. The customer's like for lamb chops12. This is an example of anticipated positive emotion (enjoy-
ment) resulting from eating/getting lamb chops
2. The customer's sustainable/equitable development values. This is part of the moral values and per-
sonal standards factor, which depends on the consumer's social identity13 (Bagozzi et al. 1996),
3. The same-week purchases of big fish. This is part of the frequency/recency of past behavior factors,
and in this case, consists of customer's transaction history.
Some of this data is specific to the customer-- for example, their transaction history or their liking of
lamb chops-- while another pertains to a category that the customer belongs to, for example that of
fair traders. Also, some of the data is specific to a product (e.g. liking lamb chops), while others are
more generic (e.g. liking any fair trade product); we talk about product variations in section 5.2.
Bagozzi's framework does not explicitly mention the consumer's socio-demographic category, al-
though some of it is implicit in the consumer's social identity. We see such data, especially the one
relating to lifecycle processes, as the main driver of needs, particularly, basic ones (food, clothing,
shelter, healthcare, education). Our imaginary scenario of section 2.1 also highlighted the role of rela-
tionships in anticipating a consumer's needs. Some of these relationships are implicit in the lifecycle
concept, namely, in the nesting stages where the consumer makes purchases (or purchasing decisions)
for the benefit, and on behalf, of dependents.
At any given point in time, a consumer may be in a particular lifecycle stage. They can go from one
stage to the next, in one of two ways, 1) after some time has elapsed, and 2) after some event has oc-
curred. For example, the typical family lifecycle has a stage for young adults with preschool-age chil-
dren, followed by a stage for school-age children. For a given child, the duration of the "young adults
with preschool-age children" stage is 4 to 5 years! Similarly, the family lifecycle has an "unattached
young adult" stage, followed by a "newly married adult, no children" stage. Marriage, or otherwise
attachment, indicates the transition from one stage to the next. To the extent that a CEM system will
typically not have access to court marriage certificates, some events will have to be inferred from
property value changes. For example, teenage or young adult children move out, they start having
basic household needs (furniture, telecom services, etc.). Such an event can be detected by the change
of address, from their parents' home address to an outside address. The reverse move back home may
also be of interest to marketers, as the disposable income of the returning children increases.
12 ignoring, for the time being, the fact that it is the customer's significant other who likes lamb chops.
13 Which may be characterized, if not defined, by the customer's membership to, or militancy within, various organizations
and associations. In our example, our shopper Chris is member of www.equiterre.org
Fig. 6 shows a customer profile metamodel. Starting with the upper half, it shows that Consumer's
have properties (Property) and have values for those properties (PropertyValue). The association
between a Consumer and a PropertyValue is characterized by a confidence level, reflecting our con-
fidence in the value. For example, a value that was entered by the consumer when signing up for a
service is more reliable than a value inferred by data mining techniques.
Fig. 6. A customer profile metamodel, taking into account decision factors, demographics, and lifecycle data
Consumer's belong to Categories, with some confidence level. For example, we may not know for
sure that Chris is a socially responsible consumer (a Category), but we know that she constantly ac-
cesses the web site of equiterre.org, and posts comments on their articles. From which we infer, with
80% confidence level, that the consumer belongs to the category of socially responsible consumers.
Category'ies are also characterized by instances of Property. The relationship between a Category
and a Property (the association class ValueRange) shows which property values are admissible, or
expected, for this category. For example, for the property PreferredTransportationMode, socially
responsible people may be known to prefer public transportation or self-propelled vehicles. value
ranges for properties of categories provide default property values for members of that category. For
example, we may have no value for the PreferredTransportationMode property for Chris, but be-
cause she is a member of the socially responsible consumer category, we can infer that she prefers
-name : string
-name : string
-min : PropertyValue
-max : PropertyValue
-restrictions : object
-value : object
-min : PropertyValue
-max : PropertyValue
-restrictions : object
-confidence-level : float
-confidence level : float
-confidence level : float
Similar to the concept of Category, we have the concept of State. A consumer can be in different
states, which correspond to qualitatively different consumption behavior, in terms of needs, attitudes,
and the like. This accounts for the family lifecycle theory, where states represent lifecycle stages. It
can also be used to represent finer concepts such as 'diaper stages', discussed in the scenario of section
2.1. StateTransition's are triggered by events (EventType), which can be property change events
(PropertyValueChangeEventType, as in getting married), or timer events (TimerEventType), as in
going from size 3 diapers to a size 4 diapers after a fixed number of months. A property value change
may be a worthwhile event, in general, or if the initial value of the property equals a specific value
(e.g. moving out of the parents home, signalling need for furniture), or if the destination value equals
a specific value (changing marital status, or moving to a significant other's house, or back to the par-
ents house), hence the association classes FromValueRange and ToValueRange from Property-
ValueChangeEventType and PropertyValue.
The Relationship association class from/to Consumer. This relationship is used to represent personal
relationships of consumers, to the extent that they are relevant to consumption behavior. In our fic-
tional shopping scenario, Chris gets lamb chops because her mate likes them. She was also told about
size 4 diapers because the system suspected that she cared for a toddler, who had been size 3 for X
months. Some relationships transfer needs: if I care for a baby, the baby needs diapers, then I need
diapers-- and will be attentive to diaper ads, for example. The message to me may be adapted to my
role in the purchasing process: whether I am an advisor, a decision maker, or a doer. We use the at-
tributes needTransfer, and purchaseRole to represent these distinctions.
Notice that the model in Fig. 6 is more of a metamodel or a high-level ontology than a customer pro-
file data model, per se. An actual customer profile model can be constructed by instantiating the
classes shown in Fig. 6. For example, we need to figure out which properties are worth representing
about a customer-- depending, on part, on the product or service being sold, which consumer catego-
ries are worth considering, which lifecycle stages are relevant to the kind of product we sell, etc. This
will be part of the framework instantiation, discussed in section 7.
5.2 Product data
We showed in the previous subsection the kind of data that a company needs about its customers to,
1) assess their needs and desires (e.g. demographic and lifecycle data), 2) understand their attitudes,
biases, emotions, and values that come into in the process of satisfying those needs. This information
is used in one of two ways: 1) identifying, those, among its products, that best help address the needs
and desires of those customers, and 2) positioning those products in a way that appeals to those cus-
tomers, considering their (known) attitudes, biases, emotions, and values. To take full advantage of
our knowledge about the consumer, we need a 'comparably rich' representation of the products and
services sold by the company.
In our grocery shopping scenario, we knew the function of our products (nourishment), their composi-
tion (sodium content, for crackers, but also mercury content, for big fish), their assortments (a Shiraz,
to go with lamb chops), and their production process (faire trade or not). Different kinds of products
will have different facets. Fig. 7 illustrates some of what our data could look like. We will comment
on the important aspects of the model.
0..** is in category
process 0..1 *
Fig. 7. An example of product representation, for rich CEM functionality
Products are represented using five facets: 1) function, which represents the 'objective utility' of the
product (e.g. nourishment, transportation, cleaning product), 2) form, which represents things related
to packaging (e.g. bulk, by unit, six-pack) and presentation (visual aspects), 3) structure, can be used
to represent the ingredients or nutritional information of a food item, or the major functional compo-
nents of a device that are worth talking about (e.g. a car having a V6 turbo engine, all-wheel drive, 17
inch wheels), 4) emotions, which represents, in a shorthand form, the 'emotional function' of the prod-
uct (the emotions it elicits), and 5) manufacture, which provides information about the manufactur-
ing/construction process of the product. As a first-cut model, we include the process itself, the juris-
diction where the product was produced, and the manufacturer, all of which being information that
environmentalists, fair traders, and various social activists might care about. A conscientious con-
sumer may refuse to purchase the product of an environmentally unfriendly process (e.g. fish captured
by bottom trawling), or manufactured in a jurisdiction known for poor environmental or labor regula-
tions, or by a manufacturer operating in such a jurisdiction, or headquartered in a jurisdiction known
as a tax haven. Finally, the association class Association represents positive (goes with) or negative
(does not go with) associations between products, and the strength of the association (goes really well
with), as in "the Bleu de Brest blue cheese goes really well with the 2011 Australian Shiraz".
In the same way that consumers are members of categories (market segments), products belong to
categories. Each product category is characterized by the same five facets (function, form, structure,
emotion, and manufacture). Categories hold default information known about classes/sets of products,
which can be overridden for specific products. The categories are organized along specialization hier-
archies, using the subcategory relationship. We can have different categorizations, based on function,
country of origin, packaging, etc. Finally, similar to the associations between products (goes
with/does not go with), we have associations between categories: red wines go well with meats, white
wines go well with fish, and polka dot shirts do not go well with pinstripe suits.
5.3 Other kinds of data
Different types of types of data may be relevant to interactions between the seller and the customer,
with the purpose of capturing the needs of a customer, their appreciation of a particular product or
product category, information about products, such as technical product reviews, product compari-
sons, as well as promotional material, as described in section 4.4.2. Good experience management
depends on soliciting (customer → seller) or presenting (seller →customer) the right information at
the right time. In this section, we illustrate the kind of representations of some of this data that would
support fine-grained matches between products, customers, and context.
Let us take the example of promotional material. We showed in section 4.4.2 how the message con-
veyed by promotional material can be matched to the customer and the situation at hand for maximum
impact. To be able to realize this match automatically, we need to "index" the promotional material
with a precise description of the message.
Figure 8 shows a sample model for representing promotional material. The model isn broken in two
parts (separated by the dashed line). The upper part reproduces a subset of the models of Fig. 6 and
Fig. 7 that embodies our 'marketing knowledge', namely:
1) the fact that certain consumer categories (ConsumerCategory) have a need for specific functions
2) those functions are supported by certain products (Product) and product categories (Category),
3) those products or product categories induce Emotion's.
Recall from section 4.4.2 that in order to re-enforce/strengthen a customer's desire of a particular
product or function (the so-called anticipated positive emotions), I can show them a person, they can
identify with, who is using the function of the product, and who is experiencing those positive emo-
tions! Hence the Actor (as in comedian in a video), who is identified as belonging to a particular
category. That video clip (or some other promotional material) would be indexed with 'the action'
taking place, which is a ProductUsage of the target Function of the target Product, and should be
shown to experience the target Emotion. Thus, for the latest smart phone, I could have two different
clips: one featuring college students exchanging using their phones social media, and another clip
showing business people/professionals using the office productivity tools of the phone. I would then
send the appropriate clip to the appropriate customer, by matching on customer category.
is in category
is in category
Description of (specific) promotional material
Figure 8. A representation of promotional material
6 Computational intelligence for customer experience management
6.1 Data mining techniques for e-commerce
Data mining techniques have long been used in electronic commerce, for a variety of usages. Some
of the first usages included the use of navigation traces for a variety of purposes, including knowing
more about incoming traffic (geographic location of users/potential customers, referring pages),
optimizing the design of websites, and recommender functionalities, again with several uses, from
suggesting pages to look at, from the pattern of traces, to suggesting products, based on customer
profiles, on purchase history, or basket market-like analyses, of the kind we find on amazon.com
(see e.g. for a survey of the early uses [Sarwar et al., 2000]).
As B2C customers started leaving textual traces on the internet, text mining techniques found many
additional customer relationship management applications, including: 1) "intelligent routing" of us-
ers queries left on service portals, 2) sentiment analysis to get a feel for how customers feel about a
company's products or actions in terms of a unidimensional polarity, and 3) opinion mining, for finer
analyses of customers' views towards its products and actions; see e.g. [Lieu, 2015] for a through
treatment of the various mining flavours.
The advent of social media has opened up a whole new set of possibilities, but also challenges:
1) On the positive side, whereas the textual traces left on customer service portals were typically
limited to actual customers after they had made purchasing decisions, companies can now find
out what is said about them and their products on other media, by non-customers, or not-yet
customers, before they make a decision, and are then capable to act upon, and influence some
of the earlier steps of the purchasing process (see e.g. [Poria et al., 2013])
2) On the negative side, the proliferation of social media poses many operational challenges:
a. CRM/CEM specialists now need to collect the information from many different sources
(specialty blogs, portals, Facebook posts, Twitter tweets, etc.)
b. We are no longer able (or less able) to trace the opinions or sentiments that are ex-
pressed on the web, to individuals that we can reach out to address their needs or
change their perceptions14.
c. Our inability to manage identities across different social media. For example, if we find
three similar complaints posted on three different media, under three different pseudo-
nyms, do we have three different service failures, or is it the same person blasting us
In the next section, we look at the different potential uses of data mining techniques within the con-
text of CA-CEM
6.2 Data mining for customer experience management
Data mining techniques can be used in three areas:
1) categorization, which refers to both the identification/specification of categories, as in iden-
tifying or characterizing a particular consumer group (e.g. DINKs, "double-income no kids"),
and the assignment of individuals (a particular consumer) to categories, e.g. recognizing that
the customer in front of me is a DINK
2) opinion/belief mining, to figure out what the customer thinks or feels about things or issues,
3) feedback assessment, which refers to the ability of sellers to automatically assess-- and act
upon-- the feedback left by customers on their product ownership experience, or on their
customer experience, i.e. their service interactions with the seller.
We discuss below the main issues raised by these three applications, and any operational implica-
tions on our CA-CEM development framework.
Categorization is a very important aspect of customer experience management: the gist of CEM is
'knowing the customer', i.e. associating them with a consumer group that has identifiable needs,
desires, and consumption patterns. As we saw in Fig. 6 of section 5.1, a lot of information about the
consumer will be inferred from our knowledge of her/his membership in specific consumer catego-
14 if an unhappy customer leaves a complaint on the company's portal, the company is able to connect with that customer
and remedy the situation.
ries, as opposed as being specifically given/known for that consumer. Categorization raises two re-
lated issues: 1) how to discover/codify categories, and 2) how to categorize/classify new instances
individuals. The two issues are related, and will be discussed jointly. We discuss below two flavors of
categorization, depending on how much prior marketing knowledge was have about the market that
Categorization through an explicit codification of marketing knowledge
We codify marketing knowledge in the form of identifiable consumer groups along with socio demo-
graphic properties/property value ranges. Those properties can be defining (for example for an age
group, age is defining) or characteristic (for example for residents of a particular neighborhood, in-
come level or ethnic origin are characteristics). Knowledge representation theory may dispute the
Although definitional properties may be difficult to obtain from customers (age, income), we can
predict them from characteristic properties directly observed, for example what brand of cheese do
they typically purchase. These predictions can be made by machine learning based algorithms
trained on customer sets for which we have categorization and characteristic properties. We can
also imagine a tool that enables us to tag the properties of a model as defining versus characteristic,
thus triggering real time analyses as data is inserted.
New instances get categorized based on whichever combination of properties are available. The
system can then guess the value of other properties from the definition of the category with differ-
ent confidence levels, which would determine the strength of the inferences that we draw from
those inferred property values.
Categorization through unsupervised learning
The idea here is to use unsupervised learning techniques to discover new consumer categories. Clus-
tering algorithms are fed customer data, which is grouped under categories. The quality of the cate-
gories will depend on the richness of the data. This is somewhat related to the earlier discussion
about definitional versus characteristic properties. Among the relevant issues is whether the data
that is being captured is discriminating for the instances that are being fed or not.
Once instance data (consumer or product) has been clustered, we can compute, for each property,
value distribution for the instances of the cluster to get a better characterization of the cluster. Thus,
if a customer is found to belong to a certain category, and is missing some property values (e.g.
yearly income), we can infer values for those properties based on the property value distribution for
that category15. For example, if I know that a customer is a DINK (double income, no kids), and I
15 In fact, the property value is/can be mapped to the property value distribution of the category
have a family yearly income distribution for DINKs, I can translate that knowledge into income level
probabilities, or fuzzy income values. This, in turn, will determine the strength of the inferences I
draw from those income values.
Clustering algorithms tend to be computationally intensive. If customer categorization functionality
is used by checkout/cash register clerks for cross-sells16, then we need to make the categorization
algorithm fairly quick. This could mean many things, including: 1) using incremental clustering algo-
rithms, or 2) make an approximate, temporary category assignment, to be refined later on in batch
mode during off or low traffic hours.
One of the major problems with automatic clustering algorithms is that the resulting clus-
ters/categories are not labeled in a way that is meaningful to a marketer. Within the context of a CA-
CEM development framework, we need to provide analysts/CA-CEM application designers with a set
of tools that can perform such computations through minimal configuration.
6.2.2 Mining beliefs and values
Section 3 showed that social identity, and its corollary, second-order moral values and self-evaluative
standards influence key purchasing decision steps, namely, goal desire, behavior desire, and imple-
mentation intention. In other words, who we are (or we think we are), and our values/what we believe
in, influence both, 1) what we desire, and 2) how we go about fulfilling that desire. Chris, our mystery
shopper of section 2.1, is a "fair trader", and as such, was told about the latest fair trade certified arri-
vals. However, how can we tell if a customer is environmentally conscious, or a fair trader, or some-
one who cares about, not only the price of products they purchase, but also the labor practices of
the companies, or jurisdictions that produce them?
There are many strategies for inferring such "opinions" or "beliefs". For example, a shopper whose
basket regularly includes items that are certified organic-- or simply picked up from the or-
ganic/natural foods section-- can be safely assumed to be, a) health conscious, and b) probably rela-
tively well-to do! Idem for fair trade coffee. As more and more food store chains carry kosher or ha-
lal sections, we can even guess consumers' religions! Thus, basket market analysis techniques can be
used if we have, a) a history of purchases, and b) appropriate product labeling.
Alternatively, we can mine consumers' opinions from the traces they leave on the web. Again, it is
difficult to devise a general strategy: it depends on the kind of opinion that is sought, and on the
traces that are available. Roughly speaking, the search for opinions can be decomposed into two sub
problems: 1) finding out if a customer cares about a particular issue (topic), and 2) finding out how
they feel about that issue. To answer the first question (topic), we can check issue-specific social
media. For example, any article or comment you enter on www.climatecentral.org shows that you
care about climate change. The same can be said if the site is listed on your Facebook page, or if you
subscribe to #ClimateChange hashtag through your Twitter account. Idem if the phrase 'climate
change' occur in your posts anywhere. Generally speaking, the more diffuse the topic of the medium,
the more sophisticated the opinion mining technique is needed.
16 For example, the cashier just scanned cheese, and they know that the consumer is a DINK, they may suggest to them a
matching bottle of wine in the mid-to-expensive price range.
Once we know that you care about a particular issue, how you feel about it becomes slighter easier,
thanks to sentiment analysis techniques. For example, to find out whether you believe in human-
induced climate change or not, I can look-up your posts on www.climatecentral.org and do a simple
polarity analysis17. However, there are cases where simpler cues can go a long way towards mining
opinions. For example, if you use Monsanto and genetically modified foods (or GMF) in the same
sentence, then your position on the subject of genetically modified foods is easy to guess.
From an operational point of view, one of the issues we need to address is whether our CA-CEM
application should keep updating users profiles incrementally, and if so, what should be the trigger-
ing event for the incremental reevaluation? should it be a new Facebook post? a nightly web crawl
for my customers? a purchase? or should I run a batch process on my customer database to com-
pute their environmental opinions?
Despite advances in semantic opinion mining (see e.g. [Cambria et al., 2013], [Poria et al., 2013],
[Poria et al., 2015], [Liu, 2015]), opinion mining remains an art, because the best-- or only--available
strategy to use in each case depends on, or greatly benefits, from, 1) the selection of sources from
which to mine such opinions (e.g. available media specialized in the topic of interest), and 2) the
exploitation of information or constraints specific to that opinion (e.g. the public image of Mon-
santo, a specific company, in the agricultural food industry). Within the context of our framework,
we can only provide a tool box that 'opinion miners' can use, and it is probably something that
should be done off-line, e.g. disconnected from the transactional databases/systems.
6.2.3 Assessing consumer feedback
The model of the purchasing process shown in Fig. 2 of section 3 showed that the purchasing proc-
ess ends with a post-purchase feedback whereby the consumer expresses, in one way or another,
whether they have attained the 'desired goal' that triggered the whole process in the first place.
Naturally, companies are very interested in what their customers have to say about, 1) the products
they purchased, and 2) the quality of the service they got during the purchasing process, where the
latter has been steadily gaining in importance in building customer satisfaction and loyalty.
There are many ways of getting this feedback. Post-service customer surveys are fairly common
nowadays, where customers are given incentives to fill out customer surveys. While such surveys are
undoubtedly very valuable because of their granularity and focus, they suffer from a number of
1) a low return rate, despite the incentives
2) they don't address image and perception issues in the general public, to the extent that they
focus on actual customers, and
17 Knowing the orientation of the site alone does not tell us whether a frequent 'commentator' shares the 'editorial line' of
the site: a regular commentator may express systematically contrarian opinions.
3) the artificially accommodating/agreeable reviews that all but the most irate consumers end up
entering, at the insistent requests of customer service personnel.
Thanks to the internet and social media, consumers are now leaving unstructured, unscripted, and
uninhibited product and service reviews all over the place. This is a very active research area (see
e.g. [Cambria et al., 2012],[Vinodhini et al., 2012], [Poria et al., 2013], [Poria et al., 2015] [Liu, 2015]).
Thanks to knowledge bases combining lexical, semantic, and sentiment/affective knowledge, recent
opinion mining techniques are able to go well beyond polarity to explore emotions and intensity and
even identify fake product reviews (see e.g. [Poria et al., 2013], [Liu, 2015]). As we saw in section 3,
the anticipated positive emotions and anticipated negative emotions are important deciding factors
in goal/behavior desirability. Emotions are not only important indicators of product appreciations,
but they are very valuable measures of the success or failures of service encounters (see e.g . [Cook
et al., 2002]).
Within the context of our framework, we need to provide resources and tools to CA-CEM applica-
tions designers that they can use to mine textual/unstructured product or service reviews left on the
seller's portal. Anything beyond the data left on the seller's own web site offers the same opportuni-
ties and challenges described in section 6.2.2.
6.3 Approximate reasoning to deal with incomplete data
As mentioned earlier, the personalisation embodied in experience management centers around, 1)
selecting the products and services appropriate for the consumer's needs and desires, and 2) presenting
them to the consumer in an appealing way. Both CEM functions rely on the company's knowledge of
its products and its customers. The models of Fig. 6 and Fig. 7 showed fairly elaborate representations
of customers and products, respectively. It is very likely that the company will not have this level of
information about its products or customers. Further, for many products and purchase types, much of
this information will not be relevant.
We outline four strategies for dealing with incomplete data:
1. Instance data versus class data: for both products and consumers, we made a distinction between
individuals (Consumer and Product) and classes/categories (Category, in both Fig. 6 and Fig. 7,
and State in Fig. 6). If we don't have information about an individual, we look it up in the cate-
gory. Thus, if I don't know that lamb chops go with the 2011 Shiraz, at least I know that meats go
with red wine, and lamb chops being meats, we know that we can recommend red wine-- but not
the 2011 Shiraz, specifically. Thus, we have 'graceful degradation' in the recommendation quality.
2. Category inheritance: if some information is missing from a category, we can look it up in its su-
per-categories. Marketers may know something about transportation needs of DINKs (double in-
come, no kids) in general, but not specifically about DINKs under thirty.
3. Because of the above, our confidence in the information about a consumer or a product depends
on the length of the inference path (from instance to category, and up in the category hierarchy).
4. Use a scoring approach, whereby the different facets contribute positively or negatively to a
<product , consumer> match. Consider two customers C1(needs: detergent, values: environment)
and C2(needs: detergent, values: none), and three detergents D1(function: detergent, manufactur-
ing.process: contains phosphate), D2(function: detergent, manufacturing.process: does not contain
phosphate), and D3(function: detergent, manufacturing.process: NA). Assume that a match on
function counts for 100, and a match on manufacturing process counts for 20 (or -20). The addi-
tive scoring approach would yield the following:
a. For customer C1, match(C1, D1) = 100 -20 = 80, match(C1, D2) = 100 + 20 = 120, and
match(C1, D3) = 100 + 0 = 100. Thus, D2 is the best match
b. For customer C2, match(C2, D1) = match(C2, D2) = match(C2, D3) = 100 + 0 = 100. In
other words, customer C2 does not care. They may use other criteria to select (e.g. cost).
Note that, when the customer does not care, the company's own values could come into play, to
position one product higher than another.
7 Instantiating the framework for a CEM scenario
Section 4 laid the foundation for CEM by outlining the range of data and functions that may be used
to support CEM functionality within the context of a generic e-business system. In this section, we
outline steps for selecting and customizing the generic functionality presented in section 4 to handle a
particular CEM scenario.
7.1 Specifying the relevant purchase scenario
In section 3, we presented the generic structure of a purchasing process, as a goal-directed behavior,
and identified the various factors that influence, or are relevant to the purchasing process. The process
of Fig. 2 may be an overkill for grabbing a carton of milk, or may not reflect the reality of impulsive
purchases where one goes window shopping, and ends up making purchases, or walks into an elec-
tronic store to buy a USB key and comes out with a laptop.
Thus, the first step in using our CA-CEM framework is to specify the purchasing scenario(s) appro-
priate for the business:
1. which steps of the process in Fig. 2 are relevant. This depends more on the type of purchase (see
e.g. (Solomon, 2009))
2. which factors are relevant to those steps. This depends more on the type of need (e.g. food versus
entertainment), and the type of product (basic versus luxury)
3. when and where do the various steps occur. This is important for many reasons: a) when, and how
much time does the consumer have to make a decision, and b) what is the best way to position the
products to the customer. This, in turn, also depends on the kind of store/products (e.g. Sharper
Image versus Home Depot), and the sale channel used (in-store vs on-line).
A given business can enact several purchase scenarios, depending on the product and on the channel,
and a CA-CEM system needs to support all the scenarios. Further, it may need to guess which type of
purchasing process is in progress, e.g. using the browsing pattern (physical or on-line) in the store.
The outcome of this step will be a set of processes, each one of which representing a subset of the
process of, with timing and location data (on-line, in-store) attached to each step.
7.2 Specifying the desired level of functionality
The model of Fig. 4 shows that all the steps of the purchasing process and all the influence factors are
amenable to CEM intervention/functionality. Based on the process specified in section 7.1, and when
and where each step takes place, a retail business needs to figure out, for each process step:
1. whether intervention at a particular process step brings value: a grocery store may not need to
intervene at the Goal desire step, as a consumer needs no enticements to buy a carton of milk.
2. whether it has the means to intervene. This is an issue of opportunity and physical means. If you
drop lamb chops in the shopping cart, and I have no way of finding that out (physical limitation)
until you reach the cashier, it is probably too late (opportunity) to suggest an appropriate wine.
3. what is the best modality. Intervening at various steps of the purchasing process can easily be-
come intrusive. Should I "push" specials to the consumer (the Australian wine special), unsolic-
ited, or wait until s/he asks for wine recommendations, or suggest a wine when they pick a prod-
uct that goes with wine?
Notice that the answers to these questions can help a brick-and-mortar retailer decide which sensors to
put, where, to get the best cost-effective CEM solution. again, contrast this approach with "stick an
RFID tag to everything that moves, and see what you can do with it".
7.3 Build the relevant data schemas
CEM functionalities rely on our knowledge of purchasing processes, products, and consumers. We
saw in sections 5.1 and 6, a) the range of information we can manipulate about consumers and prod-
ucts to support CEM, and b) how default information about products and consumers can be gleaned
from the categories to which they belong.
The desired level of functionality specified in the previous section enables us to figure out what/how
much data about consumers and products we need. We now have to, 1) specify the structure of the
data, and 2) populate it for categories (for consumers and products) and states (for consumers).
The models shown in Fig. 6 and Fig. 7 represent ontologies, i.e. metamodels of the actual data models
needed by our CA-CEM functionality. In this step, we need to instantiate those ontolo-
gies/metamodels to specify the actual data schemas that are needed by our system. For example, the
model in Fig. 6 shows that a Consumer can have a bunch of Property's, and that each Category is
characterized by a range of values for those properties. In this step, we need to specify the properties
that we are interested in. For a Consumer, such properties include your typical CRM data such age
(or birth date/birth year), address, payment methods, profession, perhaps income bracket, marital
status, and transaction history. To this, we add their significant relationships, their social identity(ies),
their values. Consumer likes for specific products or product categories can be represented by ternary
Most of these Property's may be used to characterize Category'ies (e.g. DINKs, generation Y) and
State's (married people, various family lifecycle stages), but some are exclusive to instances (e.g.
name, transaction history). When a property applies to a Category (or State), it is represented by a
range of values. For example, DINKs (double income, no kids), may have an income bracket of [100
k, ∞[, like sports cars, generation Y have an age range of [18 - 35], etc. Entering value ranges for
these properties is crucial, since learning of a consumer's membership to a particular Category or
State can help us infer a lot of information about their needs and desires. Those values embody the
marketing knowledge that the business has about its customers or potential customers.
A similar process needs to take place for the product data. We need to first specify which properties
are worth representing. Then, we populate information about product categories (see the model of
Fig. 7). That is where we specify that red wines go with meats, that big fish tend to have high mercury
contents, or that Absurdistan has poor labor laws and lax environmental regulations.
7.4 Populate instance data
The product instance data is a one-time deal for the products sold by the business. Idem for promo-
tional material (see section 5.3).The consumer instance data, however, is a work in progress. We show
below the lifecycle of a consumer record:
1. creation. Businesses learn about new or potential customers, in many ways: a) when they go
through the cash register the first time around, b) when they access anonymously the web site of
the company-- in which case they may be known by IP address, c) when they create accounts on
the company's web site, or post comments with an ID from another portal, and d) by purchasing
consumer lists. Depending on the channel, the business will have more or less information.
2. Category/state assignment. This the process through which a consumer is identified with a cate-
gory (a generation Y), or with a state/lifecycle state (young married couple in early childbearing
phase). This information can be entered explicitly (filling out a questionnaire when applying for a
store credit card), or inferred. For example, someone who buys baby formula or diapers regularly
is assumed with a high-level of certainty, to be in the early childbearing phase. This assignment
allows to infer other data.
3. Property value updates. Property values, or our confidence in their values, may be regularly up-
dated through, a) explicit entry, or b) an accumulation of evidence. For example, if someone has
bought baby formula three times in the past two weeks, are probably in childbearing stage (confi-
dence level of 70%)-- although they could just be staying with someone who is. If they keep at it
week after week, our confidence level goes up (95%). Generally speaking, consumer classifica-
tion rules need to kick-in each time information about the customer is updated.
4. State change. Recall, from the model in Fig. 6, that state changes can be triggered by two kinds
of events, a) timer events, and b) property value change events. A pre-school toddler will remain
in this state for a maximum of 5 years. A teenager who changes home address, from his/her par-
ents home to an outside address has probably moved into young adulthood. Thus, each time a
consumer property value is updated, we run statement assignment rules.
We mention the possibility that category data be updated from consumer instance data. For example,
our marketing department came up with a category (market segment) based on age an income. We
later realize that all of the consumers of that category live in a particular area. The zip code could be-
come a characteristic feature of this category, i.e. its value used as a default for consumers of the
category, but it would not be used as a criterion for membership (i.e. it is not definitional).
Customer experience management (CEM) aims at personalizing a customer’s interactions with a
company around the customer's needs and desires (Weijters et al., 2007). Researchers and e-business
visionaries have been fantasizing about the kind of personalization afforded by ubiquitous computing
(e.g. the 'internet of things'). While all the technical ingredients for such fantastic scenarios exist today
(sensor technology, middleware, mobile computing, social network analysis, data mining techniques),
there are no guidelines to help either, 1) e-business software vendors, to figure out which functional-
ities to provide in their software to support customer experience management functionalities, or 2) e-
business software clients to figure out which CEM functionalities to implement, let alone how to im-
plement them, given, a) the type of products they sell, b) the channels through which they sell those
products, and c) their software and hardware capabilities.
Consumers interact with companies to acquire products or services that satisfy their needs and desires:
a purchase (Bagozzi et al., 2007b). To personalize such interactions, we need to understand them. To
this end, we relied on studies of consumer behavior from marketing and social psychology. Such stud-
ies identified, a) the steps undertaken by a consumer in a purchase process, and b) the factors that in-
fluence their decision making in those steps (section 3). For the purposes of CEM, these steps are po-
tential interaction opportunities between the company and the consumer, and the influence factors
represent relevant data that the company could use to personalize those interaction (section 4.1). This
enabled us to identify an experience management pattern that could be applied to various steps of the
purchasing process. However, we showed that a 'blind application' of the pattern to all the steps of the
purchasing process described in section 3 would not be appropriate (section 4.2). We argued that the
application of this pattern is best considered within the context of service design (section 4.3), and
showed an example of the analyses then a service designer could make to instantiate this pattern (sec-
tion 4.4). While section 4 explored the design vocabulary for customer experiences, section 5 pre-
sented the kind of representation that is required for consumers and products to be able to match
products to consumer's needs/desires, preferences, biases, and attitudes. how to deal with incomplete
customer or product data (section 6). Given such an infrastructure, we proposed the first elements of a
methodology enabling e-business software users to, 1) specify the CEM functionalities they wish to
support (sections 7.1 and 7.2), 2) set-up the data infrastructure (section 7.3), and 4) populate the con-
sumer data (section 7.4).
This work is at an early stage. In this paper, we focused on the foundations for a methodology, for
designing and implementing context-aware customer experience management functionalities. The
proposed approach gave us a frame of reference to study the multitude of issues raised by customer
experience management functionalities, within the context of e-commerce software. We are currently
testing the theory on a specific purchasing scenario provided to us by an e-commerce suite vendor that
will enable us to, a) validate the theory, and b) implement a first cut of some of the software artifacts
of our framework.
[Ajzen & Fishbein, 1980] Ajzen, I., & Fishbein, M.: Understanding attitudes and predicting social.Behaviour.
Englewood Cliffs, NJ: Prentice-Hall (1980).
[Ajzen , 1985] Ajzen, I.: From intentions to actions: A theory of planned behavior (pp. 11-39). Springer Berlin
[Ajzen & Madden, 1986] Ajzen, I., & Madden, T. J.: Prediction of goal-directed behavior: Attitudes, intentions,
and perceived behavioral control. Journal of experimental social psychology, vol. 22, no. 5, 453-474
[Bagozzi & Warshaw, 1990 ] Bagozzi, R. P., & Warshaw, P. R.: Trying to consume. Journal of consumer re-
search, 127-140 (1990).
[Bagozzi, 1992] Bagozzi, R. P.: The self-regulation of attitudes, intentions, and behavior.Social psychology
quarterly, 178-204 (1992)
[Bagozzi et al., 1996] Bagozzi, R. P., Henderson, G., Dabholkar, P. A., & Iacobucci, D. Network analyses of
hierarchical cognitive connections between concrete and abstract goals: An application to consumer recy-
cling attitudes and behaviors (1996).
[Bagozzi & Dholakia , 2006] Bagozzi, R. P., & Dholakia, U. M.: Antecedents and purchase consequences of
customer participation in small group brand communities.International Journal of Research in Market-
ing, vol. 23, no. 1, 45-61 (2006)
[Bagozzi et al., 2007b] Richard P. Bagozzi, Zeynep Gurhan-Canliu, and Joseph R. Priester, The Social Psychol-
ogy of Consumer Behavior, Open University Press, 2007.
[Benzarti et al., 2014] Imen Benzarti, Hafedh Mili, and Abdel Obaid, "Fondements Comportementaux pour la
Gestion de l'Expérience Client," LATECE technical report, october 2014.
[Biswas & Blair, 1991] Biswas, A., & Blair, E. A.: Contextual effects of reference prices in retail advertise-
ments. The Journal of Marketing, 1-12 (1991).
[Cambria et al., 2012] E. Cambria, C. Havasi, and A. Hussain, "SenticNet 2: A Semantic and Affective Re-
source for Opinion Mining and Sentiment Analysis.," in proc of FLAIRS conference, pp. 202-207, 2012.
[Constantine, 2009] Constantine, L. (2009). Human Activity Modeling: Toward A Pragmatic Integration of
Activity Theory and Usage-Centered Design. Human-Centered Software Engineering, 27–51.
[Cook et al., 2002] Cook, L. S., Bowen, D. E., Chase, R. B., Dasu, S., Stewart, D. M., & Tansik, D. a. (2002).
Human issues in service design. Journal of Operations Management, 20(2), 159–174.
[Csikszentmihalyi, 1988] Csikszentmihalyi, M.: The flow experience and its significance for human psychology
[Fazio, 1986] Fazio, R. H.: How do attitudes guide behavior. Handbook of motivation and cognition: Founda-
tions of social behavior, 1, 204-243 (1986).
[Fishbein & Ajzen, 1975] Fishbein, M., & Ajzen, I.: Belief, attitude, intention and behavior: An introduction to
theory and research (1975).
[Hill et al., 2002] Hill, A. V., Collier, D. a., Froehle, C. M., Goodale, J. C., Metters, R. D., & Verma, R. (2002).
Research opportunities in service process design. Journal of Operations Management, 20(2), 189–202.
[Liu, 2015] B. Liu, Sentiment Analysis: Mining Opinions, Sentiments, and Emotions. Cambridge University
[Mylopoulos et al., 1999] Mylopoulos, J., Chung, L., & Yu, E. (1999). From Object-oriented to Goal-oriented
Requirements Analysis. Commun. ACM, 42(1), 31–37. http://doi.org/10.1145/291469.293165
[Ostrom et al, 2010] Ostrom, A. L., Bitner, M. J., Brown, S. W., Burkhard, K. A., Goul, M., Smith-Daniels, V.,
Rabinovich, E. (2010). Moving Forward and Making a Difference: Research Priorities for the Science of
Service. Journal of Service Research. http://doi.org/10.1177/1094670509357611
[Palmer, 2010] Palmer, A. : Customer experience management: a critical review of an emerging idea. Journal of
Services Marketing, vol. 24, no.3, 196-208 (2010).
[Patricio et al., 2011] Patricio, L., Fisk, R. P., Falcao e Cunha, J., & Constantine, L. (2011). Multilevel Service
Design: From Customer Value Constellation to Service Experience Blueprinting. Journal of Service
[Poria et al, 2013] S. Poria, A. Gelbukh, A. Hussain, N. Howard, D. Das, and S. Bandyopadhyay, "Enhanced
SenticNet with Affective Labels for Concept-based Opinion Mining," in IEEE Intelligent Systems, no. 2,
pp. 31-38, 2013.
[Poria et al., 2015] S. Poria, E. Cambria, A. Hussain, and G.-B. Huang, \Towards an Intelligent Framework for
Multimodal A_ective Data Analysis," in Neural Networks, vol. 63, pp. 104-116, 2015.
[Sarwar et al., 2000] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, \Analysis of Recommendation Algorithms
for E-Commerce," in Proceedings of the 2nd ACM conference on Electronic commerce, pp. 158-167,
[Schiffman & Wisenblit, 2014] Schiffman, Leon G. & Wisenblit, Joseph (2014), Consumer
Behavior, Prentice Hall, 528pp
[Schwartz, 1992] Schwartz, S. H.: Universals in the content and structure of values: Theoretical advances and
empirical tests in 20 countries. Advances in experimental social psychology, vol. 25, no. 1, 1-65 (1992).
[Solomon,2009] Solomon, M. R.: Consumer behavior: buying, having, and being. Pearson (2009)
[Svee & Zdrakovic, 2015] Eric-Oluf Svee and Jelena Zdravkovic, "Case-based Development of a Taxonomy of
Consumer Preferences Demonstrated Using Brand Personality", to appear in the proceedings of Principles
of Enterprise Modeling (PoEM) 2015.
[Teixeira et al., 2012] Teixeira, J., Patrício, L., Nunes, N. J., Nóbrega, L., Fisk, R. P., & Constantine, L. (2012).
Customer experience modeling: from customer experience to service design. Journal of Service
Management, 23(3), 362–376. http://doi.org/10.1108/09564231211248453
[Verhoef et al., 2009] Verhoef, P. C., Lemon, K. N., Parasuraman, A., Roggeveen, A., Tsiros, M., &
Schlesinger, L. A.: Customer experience creation: Determinants, dynamics and management
strategies. Journal of Retailing, vol. 85, no. 1, 31-41 (2009).
[Vinodhini & Chandrasekaran, 2012] G. Vinodhini and R.M. Chandrasekaran, "Sentiment Analysis and Opinion
Mining: a Survey," in International Journal of Advanced Research in Computer Science and Software
Engineering, vol 2, no 6, pp. 282-292, 2012.
[Weijters et al., 2007] Weijters, B., Rangarajan, D., Falk, T., & Schillewaert, N.: Determinants and outcomes of
customers' use of self-service technology in a retail setting. Journal of Service Research, 10(1), 3-21
[Zdravkovic et al., 2013] Jelena Zdravkovic , Eric-Oluf Svee and C. Giannoulis, "Capturing Consumer
Preferences as Requirements for Software Product Lines", in Requirements Engineering, 2014, pp. 71-90.