message perception within context-aware recommender systems

15
Message Perception within Context-Aware Recommender Systems Mark A. Hooper, Paul Sant University of Bedfordshire, Department of Computer Science and Technology, University Square, Luton, UK, LU1 3JU [email protected], [email protected] ABSTRACT Thus far implementation of Context Aware Recommender Systems have primarily focused on what to recommend by deriving results from patterns of behavior and environment to determine optimum product selection for recommendation. Our experiment demonstrates that a purchaser’s affective state also has an effect on their perception of information presented via a mobile device. We posit that the ‘how’ and ‘when’ to recommend are important considerations that have not been fully addressed when considering the display of recommendations. Together with user behaviors associated with purchasing traits, e.g. impulse buying, we explore the information processing styles of mental imagery and analytical processing; risk acceptance; involved user effort; and marketing techniques of positive and negative appeals. Results show that these different methods of presenting information to the purchaser will be successful in obtaining a positive user perception within different affective states. Together an understanding of these information presentation and processing techniques is used to build a representation of a purchaser’s perception that could be used in m-commerce systems. KEYWORDS Recommender systems, personalization, user interfaces, affective computing, context-aware 1 INTRODUCTION Research is beginning establish an under- standing of user affective, social and phys- ical states and their relevance within con- text-aware systems [1]. However it is only now with the advance of smart-phone sen- sor technology that research can truly lev- erage this knowledge within the area of mobile recommender systems [2]. Though research into context-aware recommender systems is now showing positive results through multi-criteria evaluation of both user generated content and environmental context the utilisation of contextual infor- mation is still thus far limited. The focus of this paper is to demonstrate that user context can be used to understand how an individual reacts to information presentation styles via a mobile device. We posit that understanding user behavior within context is critical to fully realise the potential for recommender system results through message customisation, especially within the developing area of m-commerce environments. To support this we define and partly verify a framework for recom- mender system personalisation that intro- duces a new layer of system intelligence through the use of message customisation based on user contextual behavior. We discuss the theory that mood and emo- tions influence our selection of cognitive processing modes which in turn provide an Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015 ISBN: 978-1-941968-08-6 ©2015 SDIWC 59

Upload: liezelleann

Post on 17-Jan-2016

232 views

Category:

Documents


0 download

DESCRIPTION

Thus far implementation of Context Aware Recommender Systems have primarily focused on what to recommend by deriving results from patterns of behavior and environment to determine optimum product selection for recommendation. Our experiment demonstrates that a purchaser’s affective state also has an effect on their perception of information presented via a mobile device. We posit that the ‘how’ and ‘when’ to recommend are important considerations that have not been fully addressed when considering the display of recommendations. Together with user behaviors associated with purchasing traits, e.g. impulse buying, we explore the information processing styles of mental imagery and analytical processing; risk acceptance; involved user effort; and marketing techniques of positive and negative appeals. Results show that these different methods of presenting information to the purchaser will be successful in obtaining a positive user perception within different affective states. Together an understanding of these information presentation and processing techniques is used to build a representation of a purchaser’s perception that could be used in m-commerce systems.

TRANSCRIPT

Page 1: Message Perception within Context-Aware Recommender Systems

Message Perception within Context-Aware Recommender Systems

Mark A. Hooper, Paul Sant

University of Bedfordshire,

Department of Computer Science and Technology,

University Square, Luton, UK, LU1 3JU

[email protected], [email protected]

ABSTRACT

Thus far implementation of Context Aware

Recommender Systems have primarily focused

on what to recommend by deriving results

from patterns of behavior and environment to

determine optimum product selection for

recommendation. Our experiment

demonstrates that a purchaser’s affective state

also has an effect on their perception of

information presented via a mobile device. We

posit that the ‘how’ and ‘when’ to recommend

are important considerations that have not been

fully addressed when considering the display

of recommendations. Together with user

behaviors associated with purchasing traits,

e.g. impulse buying, we explore the

information processing styles of mental

imagery and analytical processing; risk

acceptance; involved user effort; and

marketing techniques of positive and negative

appeals. Results show that these different

methods of presenting information to the

purchaser will be successful in obtaining a

positive user perception within different

affective states. Together an understanding of

these information presentation and processing

techniques is used to build a representation of

a purchaser’s perception that could be used in

m-commerce systems.

KEYWORDS

Recommender systems, personalization, user

interfaces, affective computing, context-aware

1 INTRODUCTION

Research is beginning establish an under-

standing of user affective, social and phys-

ical states and their relevance within con-

text-aware systems [1]. However it is only

now with the advance of smart-phone sen-

sor technology that research can truly lev-

erage this knowledge within the area of

mobile recommender systems [2]. Though

research into context-aware recommender

systems is now showing positive results

through multi-criteria evaluation of both

user generated content and environmental

context the utilisation of contextual infor-

mation is still thus far limited.

The focus of this paper is to demonstrate

that user context can be used to understand

how an individual reacts to information

presentation styles via a mobile device. We

posit that understanding user behavior

within context is critical to fully realise the

potential for recommender system results

through message customisation, especially

within the developing area of m-commerce

environments. To support this we define

and partly verify a framework for recom-

mender system personalisation that intro-

duces a new layer of system intelligence

through the use of message customisation

based on user contextual behavior.

We discuss the theory that mood and emo-

tions influence our selection of cognitive

processing modes which in turn provide an

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 59

Page 2: Message Perception within Context-Aware Recommender Systems

insight into the level of message persua-

sion. To develop our perception trait model

we have developed hypotheses that focus

on the relationships between affective

states, cognitive capacity and behavior. It

is generally agreed that positive moods re-

sult in reduced capacity and therefore a

favouring towards heuristic processing,

whereas negative moods can facilitate

more complex detail analysis [3].

Different affective states can also influence

different purchaser traits including, moti-

vation [4], [5], impulse buying [6], com-

pulsive buying [7], brand attitude and ad-

claim recall [8], risk-taking and self-image

[9]. Myers and Sar [8] provide valuable

insight into how a pre-existing mood af-

fects a user’s response to imagery inducing

advertisements. We show that understand-

ing these cognitive ability and behaviors

should strengthen recommendation con-

version when coupled with standard rec-

ommender techniques.

The rest of this paper is structured as fol-

lows. We investigate a number of affect

behavior relationships and their affect user

perception in section 2. We then discuss

our implementation of an Android applica-

tion used to capture ‘in the wild’ user per-

ception of specific messaging styles, see

section 3. In section 4 we present and ana-

lyse our results and in section 5 we discuss

limitations and opportunities for further

research. Section 6 presents our final con-

clusions.

2 AFFECTIVE PURCHASING

BEHAVIOR

2.1 Consumer Behavior and

Advertisement Techniques

We hypothesise that understanding behav-

ior towards a set of situational contexts can

be utilised to optimise context-aware sys-

tems by providing a reasoned reaction to-

wards, not only the presented options, but

also the method of presentation to the user.

We stipulate that the addition of affective

phenomena to the contextual picture is to

also consider the user’s behavior as reac-

tional and not just as an additional element

of the context that influences preferences.

We can thus potentially indicate behavior

towards the advertisement content and the

medium (i.e. text, image or video), as dis-

cussed in the paper by [10]. This hypothe-

sis leads us to consider behavior as a key

concept to advance research within Con-

text-Aware Recommender Systems (CARS),

thus providing further potential for solu-

tions to commercial recommender system

that operate in complex environments tar-

geting audiences with distinct catalogue

product types numbering in their millions.

Though an everyday occurrence the act of

purchasing an item, whether in store or on-

line, is a complex process that includes

both environmental factors and consumer

characteristics, marketing and environment

stimuli, motivation and personality factors.

There are many drivers that form an indi-

vidual’s approach to the purchasing cycle.

These complex emotional drivers include

social potency and closeness, stress reac-

tion, control, harm avoidance, traditional-

ism, and absorption [6], enjoyment [11],

and perception of risk [12]. These in turn

influence purchasing behaviors of impulse

[6], need for convenience and information

search [13]. Personality traits generally

form our emotional responses to situations

so are key to understanding particular pur-

chasing behaviors such as impulse buying

[6]. As its name implies, impulse buying is

an unplanned event that is made through a

‘snap’ judgment process. By reviewing

stimuli to form a quick, convenient repre-

sentation of a situation it is often character-

ized as a type of holistic processing that

has advantages of speed, and reduced cog-

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 60

Page 3: Message Perception within Context-Aware Recommender Systems

nitive effort [14].

A typical example of a holistic processing

technique is mental imagery, this is an in-

fluential tool for advertisers for enhancing

brand attitudes while engaging consumers

[8]. The process not only includes the mar-

keting message cues of visual, auditory,

tactile and emotional [15], but also draws

upon the purchasers previous experience,

memories and daydreams to fully form a

visual image of the situation [13]. This

contrasts with analytical processing which

forms a comprehensive understanding of a

situation through analysis of individual

stimulus characteristics. Burroughs [14],

determines that the style of processing is

selected depending on the characteristics

of task, stimulus and the individual con-

sumer.

An individual’s purchase behavior can be

predicted through their perception of risk,

a consumer will avoid impulse buying

when perception of risk is high [16].

Bhatnagar et al. [12] report on relation-

ships between risk, convenience and on-

line shopping stating that certain product

categories. Music and CD’s, are not gener-

ally considered risky because of the practi-

calities of shopping on-line, i.e. reduction

of costs and an increase in convenience to

make purchases more likely [12]. Products

with higher value are perceived as to have

a higher risk, however they could be

viewed as being more convenient to be

purchased on-line if more involved [12], or

are likely to require an evaluation process

or other pre-purchase activity [13].

Evaluation processes used in information

search rely upon analytical information

processing to produce a comprehensive

understanding [14]. Information search via

the use of mobile phones is important in

the evaluation of alternatives and pre-

purchasing activities, e.g. finding discount

vouchers [13]. Using an analytical pro-

cessing style the individual will attempt to

understand details of the purchasing situa-

tion from all angles, in doing so they will

be more likely to identify all important in-

formation including negative factors and

therefore be able to limit risky conse-

quences [14].

In addition to considering styles of infor-

mation processing, risk acceptance and

levels of processing effort we should also

understand how common techniques for

manipulating emotions are important in

marketing campaigns. We have briefly

mentioned emotional drivers that shape our

decisions and behavior, the use of emo-

tional appeals in marketing create a psy-

chological reaction that could be resolved

by acting upon the appeal message, e.g.

through purchasing an item [17]. Fear ap-

peal has been widely used in commerce

and awareness campaigns with varied suc-

cess depending on content and severity of

message [17], however the basic premise is

to focus on insecurity and concerns in or-

der to prompt action. Positive appeals also

exist and are written to engage arouse

emotions like love, desire or humour to

invoke behaviors including self-esteem

[18]. So we can summarize the above by

identifying four categories that help form

knowledge of consumer engagement with

marketing messages and around which we

can build our hypotheses:

Processing style – mental imagery vs.

analytical

Risk acceptance – low risk vs. high risk

Cognitive capacity – low effort vs. high effort

Appeal type – positive vs. negative appeals

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 61

Page 4: Message Perception within Context-Aware Recommender Systems

2.2 The Influence of Emotion upon

Consumer Behavior

This section discusses a number of hypoth-

eses that together will represent a broad

understanding of user perception (and thus

potential user behavior) for use with

CARS. We expect that by determining a

user’s affective state as being positive we

will be able to establish a different set of

likely behaviors when compared to a nega-

tive state. This notion would then support a

systems’ approach in presenting certain

information or taking a specific action.

Myers and Sar [8] discuss the relevance of

mood and its likelihood as a context for an

advertisement to be successful. Alongside

previous research efforts they state that

their findings appear to show that positive

evaluation of an advert is enhanced when

in a positive mood through the increased

ability to undertake mental imagery pro-

cessing. They also suggest that capacity to

evaluate detailed information is reduced

during periods of positive mood but this

then increases during periods of negative

mood. This is supported by Escalas [19],

who notes that the effort in generating the

mental imagery decreases the ability to un-

dertake further cognitive tasks such as crit-

ically analyse the adverts’ content which

could in turn produce more negative evalu-

ations.

These findings suggest that mood is a use-

ful context when ascertaining how to pre-

sent items via a recommender system. The

use of mental imagery may act to make the

recommendation more appealing as mood

positivity increases and thus conducive to

the actual success of the advert. Where a

negative mood is present and mental im-

agery deemed less favourable then recom-

mender messages that provide detail suited

to analytical processing could be more

successful. Presenting recommender items

through the use of mental imagery or ana-

lytical processing depending on the user’s

affective context are a novel concepts,

therefore we posit that:

H1: Processing Style

H1a: that a positive correlation will be

achieved between user affective state and

the perception of a mental imagery induc-

ing statement

H1b: that a negative correlation will be

achieved between user affective state and

the perception of a statement using analyt-

ical, detail-oriented reasoning

The relationship between risk-taking and

mood holds a similar theme. Research has

often reported that when we are in a posi-

tive mood and are presented with a hypo-

thetical situation we are more risk favoura-

ble. For example Yuen and Lee [20], note

that those in a positive mood are less con-

servative and more open to risk. However

they do report significant differences of the

effect of mood on levels of risk ac-

ceptance. This could be explained by not-

ing Isen [21], who suggests that when a

person is presented with a real risk situa-

tion they are more likely to be risk adverse.

Therefore, along with other research such

as [22] we postulate that negative mood is

more complex than basic categories of la-

boratory induced moods of ‘sad’ as used

by [20]. In addition to this it may also be

logical to suggest that real life situational

mood and emotions may potentially pro-

duce different results to laboratory find-

ings, especially under different situational

contexts.

Previous research has also determined the

effect of mood on our perception of risk.

Lee [16], presents results that demonstrate

that elements of positive mood are related

to impulsive buying traits. Brave and Nass

[9] state that it is expected that we will en-

deavour to maintain a positive sensation by

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 62

Page 5: Message Perception within Context-Aware Recommender Systems

being more risk adverse with engagement

likely to continue with low-risk impulse

sales. In addition to this, when in a nega-

tive state we are generally aiming to recap-

ture a more positive outlook and are more

likely to engage with riskier purchases to

kick-start the positive emotional process

[9]. We follow this reasoning for the next

hypothesis pair.

H2: Risk Acceptance

H2a: that a positive correlation will be

achieved between user affective state and

the perception of a statement with a low

risk focus

H2b: that a negative correlation will be

achieved between user affective state and

the perception of a statement with a high

risk focus

An interesting consideration for the use of

mobile devices is the presentation of in-

formation in an accessible, intelligent

manner. The size of device and our general

preference for convenience should influ-

ence the way information is presented.

Large sections of text may be off putting to

a user, or indeed be preferred, depending

on their mood or other situational context.

Martin [3] reviewed several research ef-

forts. He summarises that happy moods

lean towards a shallower, heuristic pro-

cessing due to a lowered cognitive capacity

e.g. probably through being distracted or

when in a pleasant environment. Whereas

sad moods suffer more effortful processing

potentially due to a more problematic envi-

ronment. With this research in mind we

explore the following hypothesis.

H3: Cognitive Capacity

H3a: that a positive correlation will be

achieved between user affective state and

the perception of a statement with low ef-

fort processing

H3b: that a negative correlation will be

achieved between user affective state and

the perception of a statement with more

effortful processing

The use of positive and negative messag-

ing in adverts and other action appeals

have been widely discussed and used in

both industry and public sector for dec-

ades. Simple optimistic appeals to traits,

such as self-esteem [18], are commonplace

and provide positive messages to encour-

age actions that will produce a positive

outcome. Fear appeals are a different tech-

nique which are more complex and require

a greater understanding of how negative

thoughts are transferred to the user in order

to promote an action [23]. A prime exam-

ple of a use for fear appeal is the health

awareness warning focussing on long term

change, see review by [24].

Mood has also been shown to have an ef-

fect on both positive and fear appeals. We-

gener [25] observed that someone in a pos-

itive mood would be persuaded more by a

positively framed message than would a

person in a negative mood. They also

found that the opposite occurred for nega-

tively worded messages, with those in a

negative mood being more susceptible to

fear appeals. We adopt Wegener et al [25],

findings to support our next pair of hy-

potheses.

H4: Appeal Type

H4a: that a positive correlation will be

achieved between user affective state and

the perception of an optimistic appeal

statement

H4b: that a negative correlation will be

achieved between user affective state and

the perception of a fear appeal statement

The above hypotheses will enable us to

first investigate previous claims that mood

affects both cognitive capacity and user

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 63

Page 6: Message Perception within Context-Aware Recommender Systems

behavior and then develop a generic model

for message style manipulation for rec-

ommender systems and other platforms i.e.

on-line advertising. This will then, in theo-

ry, provide a platform on which to build,

with following research focusing in on

specific purchasing traits which could be

manipulated using bespoke engagement

strategies.

2.3 A Measurement of Perception

To conduct our experiment we require a

simple yet effective measure of user per-

ception. The following section discusses

the rationale for a three point questionnaire

to gauge user perception of specific mes-

sage styles in order to gauge likelihood of

engagement and conversion from browser

to purchaser.

Liu et al. [26] observe that researchers and

practitioners must understand consumer

perception in order to be effective in the

area of mobile advertising. Research into

context has partly enabled this; however,

understanding of mobile user context is

still incomplete and is a popular area of

research. User perception is fundamental to

understanding a user’s attitude towards the

advert (Aad) and thus the likelihood of an

adverts’ success [27]. Aad is the positive or

negative feelings towards an advertise-

ment, service or product within a particular

context and has a strong impact on pur-

chasing [28]. Concepts of perception and

its measurement are complex [29] and with

the measurement of Aad there is also uncer-

tainty. Even early work ascertained that

Aad, as a mediating casual variable that in-

fluences purchase intension, could follow

many possibilities [27]. Research attempt-

ing to measure Aad has produced multiple

measurement scales, Bruner [30] identifies

75 multi-item measures involving 53 dif-

ferent semantic differentials and conserva-

tively suggests an openness within the

field towards measures of Aad. Drawing

upon simple measures used to calculate Aad

to in turn imply perception we opt for a

very short and broad measure of attitude to

gauge opinion. After analyzing the differ-

ent semantic differentials presented by [30]

we reject examples that were deemed to be

more applicable to the specific content that

an advert likely to present e.g. informative-

informative and beautiful-ugly. We select

three very general semantic differentials

that capture a measure of formed attitude

and also encapsulate the majority of those

shown to have been used previously. These

are effective-ineffective, appealing-

unappealing and believable-unbelievable.

Though we are utilizing the term effective-

ness as part of our perception measure its

use should not be compared with the term

advertising effectiveness. Advertising ef-

fectiveness is widely understood to be the

final measure of an advert where the con-

sumer actually makes a purchase. Research

in this area spans decades e.g. [31], [32].

Attitude towards an advert or recommen-

dation is clearly key towards its success

[32], thus our use of effective-ineffective is

to capture a general measure of the sub-

jects’ personal perception of whether a

message is capable of producing a deep

impression or achieving its intended result.

In addition to this our use of believable-

unbelievable is to represent the user’s per-

ception of the message’s credibility, which

Lutz et al., [27] identify as a determinant

of advert attitude. The recent world-wide

study by [33] reports that credibility is a

fundamental component towards advert

effectiveness. This is consistent with the

wider research community’s opinion of the

impact of trust of mobile advertisements

and recommender systems [26].

With regards to differential appealing-

unappealing appeal is obviously a personal

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 64

Page 7: Message Perception within Context-Aware Recommender Systems

thing. It is clear that visually appealing ad-

verts accelerate a consumer’s intention to

purchase [34], however other content is

also important. Park et al. [35] determine

that appeal to emotions will be particularly

appropriate to advertisements within the

mobile environment. Hadija et al. [28] note

that users of social media do notice em-

bedded advertisement’s but quickly disre-

gard them to focus on other content such as

friend’s profiles and pictures. It is clear

that this is mostly due to the user’s focus

on the ‘task in hand’ but is also identifies

that an advertisement must focus on char-

acteristics of attractiveness, design and use

of colour to be successfully appealing [28].

Therefore we posit that a purchaser’s per-

ception of an object’s appropriateness and

design, i.e. appeal, is critical to understand-

ing likelihood of engagement and success

in its objectives.

Though simple, our three semantic differ-

entials provide an aggregate of key ele-

ments to retrieve a realistic understanding

of user message perception within real-life

situations using mobile devices. Its use is

discussed in the following section.

3 METHOD

We were keen to not follow other research-

er methodology of stimulating mood states

through techniques such as the use of

mood eliciting video within a laboratory

environment , e.g. [8], and favoured utili-

zation of natural ‘in the wild’ moods and

emotions within a mobile device context.

To complete the experiment needed to cor-

roborate our hypotheses required a robust

smart-phone application (Android) that

was user friendly, able to package data se-

curely whilst also ensuring relative unob-

trusiveness.

The following sections describes the ap-

proach to substantiate the trait behaviors

discussed earlier.

3.1 User Interaction

The main aim of the experiment was to

capture user affective state, i.e. emotions

and/or mood and measure user perception

feedback to specific statements crafted to

prove our hypotheses.

Though there have been some successes in

using sensors and other data to establish a

subject’s mood or emotions see [36], we

determined that a user’s ability to self-

diagnose (emotional intelligence) is still

more reliable and easier to implement via

the mobile phone for this stage of our re-

search. Many measures of emotion and

mood have been explored and theories

abound. However due to their relative sim-

plicity dimensional theories tend to be the

favoured approach where users are asked

to perform a self-diagnosis. We adopt

Mehrabian’s [37] Pleasure-displeasure,

Arousal-nonarousal, Dominance-

submissiveness (PAD) as it is a dominant

dimensional model which has been shown

as an effective method of modelling emo-

tions and other affective states [38]. Our

hypotheses require a scale of positive-

negative affect which directly correlates

with the pleasure-displeasure scale of PAD

[37]. In addition to this the three dimen-

sional approach provides additional granu-

larity against which to further analyse our

results and if possible determine more in-

depth hypothesis.

We poll the user during periods of natural

device use to capture the user’s self-

reporting of their affective state using a

popular psychological tool developed by

[39] called the Self-Assessment Manikin

(SAM). This three factor graphical scale

provides a quickly understood, effective

user interface, which directly transfer to

the three dimensional PAD scales. Note

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 65

Page 8: Message Perception within Context-Aware Recommender Systems

that each scale is measured from one up to

five with five being the maximum value.

Fig. 1 is a screenshot of our Android im-

plementation showing the three SAM

scales with a larger image that pops up up-

on selection of an individual image for

clearer visibility on small screens.

3.2 Determining Likelihood of

Engagement

For testing our four trait hypotheses we

developed 38 statements split over the

types of high and low risk, fear and opti-

mistic appeals, mental imagery and detail

processing, high and low effort. The long-

est statement was 343 characters long

(with spaces) with the average being 136

characters long. We wanted to ensure that

any emotion educing bias was reduced by

not using video, images or any other colour

variation throughout our tests. Therefore

all statements are displayed using the same

simple interface with light blue text on

black background using 18 px Arial. A se-

lection of five statements are randomly se-

lected and displayed, the user is then

prompted to rate each statement as de-

scribed below.

Where previous research including [40],

[41], [42], have provided comprehensive

checklists that capture user attitude to-

wards advertisements we deem that for ‘in

the wild’ testing these are too detailed to

implement. In addition, as we were not as-

sessing adverts per se but rather tailored

statements that singularly focus on a par-

ticular trait we deemed attributes such as

brand reinforcement, empathy, familiarity,

entertainment, in formativeness or state-

ment such as ‘I don’t like it’ unsuitable for

our needs. We therefore opt for a broad

three point questionnaire that focuses on

key elements of perception that have been

shown to be important factors in the suc-

cess of an advertisement, see section 2.3.

We use three semantic differentials of ef-

fective-ineffective, appealing-unappealing

and believable-unbelievable in our meas-

ure of perception. Throughout the rest of

this paper we will refer to this measure as

the eab-perception. Each message state-

ment is subject to the eab-perception, each

differential is rated using 5 point psycho-

metric Likert scale e.g. from 1) Very inef-

fective to 5) Very effective. The three eab-

perception semantic differentials have been

selected as they provide a reflective sum-

mary of the user’s perception of the state-

ment and therefore a likelihood of en-

gagement. This method enables us to pro-

duce a potential negative or positive re-

sponse to a statement. As per typical use of

multi-item we simply use the result of

SUM(eab-perception), Cronbach α = 0.79.

Each statement type is relatively simple in

structure with the focus being on a basic

message to the user for interpretation. The

structure for each type of statement used is

described as follows:

Fig.1 Screenshot of our Android implementation

showing the three SAM scales for capturing the

user’s affective state. For clarity on smaller

screens the larger overlaying image is a ‘pop-up’

which actions on selection of a smaller image

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 66

Page 9: Message Perception within Context-Aware Recommender Systems

Mental imagery – statement that only

induces mental imagery i.e. analytical pro-

cessing not required

Detail processing – statement that provides information that needs to be ana-

lysed or compared

Low effort – easy to read statement with shorter word length (18 word aver-

age)

High effort – statement with longer

word length (34 word average) using more

complex statements

Low risk – statement that presents no risk and is very general

High risk – statement that presents a situation with an element of risk

Fear appeal – statement that infers that

inaction will lead to a negative result

Optimistic appeal – statement that in-fers that an action will lead to a positive

result

The reader should note that some of the

statements are structured in a way that they

fit several categories. For example a small

number of high risk and high effort fitted

the definition of a detail processing state-

ment. The reverse however is not always

the case.

4 RESULTS

Our research into the effect of mood and

our perception of different ways of pre-

senting information via a mobile device

draws upon a number of previous hypothe-

ses. Correlation analysis is used to prove

the extent to which ‘in the wild’ affective

state determines the likelihood of each

processing behavior trait being favoured.

Our experiment collected 57 responses

with 58% of responses completed by male

participants. A total of sixteen users partic-

ipated in the experiment, 75% male and

25% female. Though a wide ethnic popula-

tion was included 62% were white British.

The spread of age groups were as follows,

21 years and under (25%), greater than 22

years and less than 35 years (50%), greater

than 35 years (25%). Though we seek

strong correlation for our results we obvi-

ously do not expect perfect values of 1 and

-1 for respective positive and negative cor-

relations for a number of reasons. We

acknowledge the main issues behind the

self-reporting technique used in our data

collection. Firstly the users’ input is sub-

jective, relying upon the user’s emotional

intelligence (ability to self-assess their

mood and emotions). Secondly that both

the input for the PAD assessment and the

perception feedback both utilize the Likert

style scales which can be prone to central

tendency bias.

As our hypotheses are biased towards par-

ticular positive or negative correlations we

test for one-tail correlation the results of

which are represented as r. The probabili-

ties of these are measured using p-values

and where statistically significant are

shown as p<0.05 (confidence level of

95%) or p<0.01 (confidence level of 99%).

Our results are as follows. We find a posi-

tive correlation between level of affect and

the perception of mental imagery state-

ments (r=0.45, p<0.01) thus H1a is prov-

en. However it is not clear that the oppo-

site case applies i.e. a negative correlation

between affect and the perception of ana-

lytical detail processing. Our results show

a minor, insignificant correlation, thus H1b

is not proven. See Fig. 2 for representative

correlations.

The failure of H1b could be explained by

suggesting that though detail processing is

favoured when in a negative state it may

not be true that we are unable to undertake

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 67

Page 10: Message Perception within Context-Aware Recommender Systems

analytical processing when in a positive

state. Though the undertaking of mental

imagery may ‘take up’ cognitive capacity

and thus reduce detail processing it does

not necessarily mean that there is an inabil-

ity to conduct some detailed analytical

processing in the absence of mental image-

ry processing. In other words though we

may not completely focus upon the mes-

sage details we will still be able to process

the overall meaning and develop a general

perception of the message [43], and there-

fore produce different results depending on

the situation.

Fig. 2 Correlations for statements using Mental

Imagery and Detail Processing

For H2a we find a positive correlation be-

tween level of affect and the perception of

low risk statements, therefore H2a is prov-

en (r=0.3, p<0.05). However the results

present no significant correlation between

affect and the perception of high risk

statements, thus H2b is not proven. To

help explain this lack of correlation we

note that Lewis et al. [44] present further

levels of complexity into the understanding

of negative emotions. He suggests that var-

ied effects on risk taking can be found with

different types of negative emotions. Con-

flicting effects are not only present when

comparing consequential with reflective

mechanisms but also that anger and fear,

while both negative, have opposing effects

with angry people favouring risk and fear-

ful people being more risk adverse.

Though this argument may provide some

insight we can also see that other context

will have an effect on mobile device user’s

perception of risk. It has been shown that

environmental factors, including sound

[45], [46] and location [47], have a con-

trolling effect on our ability to process in-

formation. In addition to this user activity

such as multi-tasking can also have a det-

rimental effect on comprehension [48].

With both environment and activity having

an impact on user ability to process infor-

mation we therefore suggest that a user

could fail to form a satisfactory perception

of risk under certain circumstances, espe-

cially if the message requires careful delib-

eration.

Our hypotheses on perceptions of effort

follow a similar pattern to H1 and H2. A

relatively low positive correlation (r=0.28,

p<0.05) has been found between user af-

fective state and the perception of a state-

ment with low effort processing. However

no useful correlation is present between

user affective state and the perception of a

statement with more effortful processing.

Therefore H3a is proven and H3b is not.

As previously mentioned [8], research has

suggested that those in a negative emotion-

al state would be more favourable towards

processing effort. However an explanation

for this suggests that this could be due to a

problematic situation [3]. Knowledge of a

user’s situation could be key to under-

standing the lack of correlation in hypothe-

sis H3b. Not only is situational context

very likely to affect cognitive ability and

willingness to engage, the limitations of

the mobile device is also likely to be a fac-

tor in preventing a user from fully engag-

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 68

Page 11: Message Perception within Context-Aware Recommender Systems

ing in effortful information processing

[49].

The pair of hypotheses involving both op-

timistic and fear appeals are also only part-

ly validated. The expected positive correla-

tion between user affective state and the

perception of an optimistic appeal state-

ment was successfully achieved (r=0.45,

p<0.01). The results for the perception of

fear appeal statements which as expected

produced a negative correlation was reject-

ed with no significance found. See Fig. 3

for representative correlations.

Fig. 3 Correlations for statements using Positive

Appeal and Fear Appeal

The value of these findings is important in

that if a recommender system is armed

with values for affective state then it can

determine how best to present the recom-

mended item. For example irrespective of

the product being recommended when the

user is in a positive state the system could

use mental imagery to maximize purchase

conversion rates and when negative use

different techniques such as increasing

brand awareness or product comparisons.

Understanding how a user’s mood shapes

their response to risk also potentially ena-

bles a system to determine how and when

to present certain higher risk items, for ex-

ample an expensive holiday. The same un-

derstanding can also be applied for levels

of effort and likelihood of engagement.

5 DISCUSSION

Whilst our ‘in the wild’ experiment has

confirmed that correlations are present be-

tween some behavior traits and levels of

affect the results show that not all cases are

proven. We reject hypotheses that require

complex analytical effort, fear appeal or

acceptance of higher risk. Concerns would

be raised if converse correlations were

achieved however we see zero (or very

close to) correlations. The spread of re-

sults, in particular for negative mood, sug-

gests that either some emotions are more

complex than tested for or that other fac-

tors are influencing the relationship be-

tween mood and the behavior traits. This

highlights the main limitation within our

experiment and shows that a larger set of

results is need for analysis against addi-

tional user contexts. In addition to this

more realistic ‘adverts’ or recommender

explanations and items important to the

user are needed to further prove the hy-

potheses developed.

While not all hypotheses have been proven

we can at this stage still present an insight

for recommender message personalization

to users in both positive and negative

moods even without considering other con-

texts. The basic behavior trait model utiliz-

es logic for personalization for users in

positive or negative affective states, see

Fig. 4. The caveat of consider_using

shown in the logic allows the option to uti-

lize different engagement techniques when

the user is not in a positive mood. While

our results from H1, H2, H3 and H4 do not

prove that techniques such as detail pro-

cessing will always be effective when the

user is in a negative state they do show that

the opposite technique, in this case mental

imagery, will not be effective. Therefore,

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 69

Page 12: Message Perception within Context-Aware Recommender Systems

selecting the consider_using options poten-

tially provides a higher chance of success.

Obviously a system that is achieving in-

sight into a purchaser’s affective state by

means other than self-diagnosis will also

have access to other contextual infor-

mation. Therefore a system will be unlike-

ly to solely utilize basic message personal-

ization described in Fig. 4 and will also

draw upon additional factors that inform

understanding of purchaser perception and

likelihood of engagement. The user’s envi-

ronment, their levels of activity and the

company of others will all have an effect

on ability to conduct actions of processing

information (whether consciously or not).

If one or more of the inputting context be-

come extreme then the purchaser may be

unable to filter out their effect causing the

perception likelihood to become less pre-

dictable. Under these situations the system

may choose to only engage in the simplest

manner, say to just increase brand aware-

ness, or indeed decline to engage com-

pletely.

Fig. 4 Pseudo code representing logic for selecting

basic message personalization.

So, to be truly successful in their integra-

tion into m-commerce recommender sys-

tems the personalized, perception-aware

interface must incorporate a range of clear-

ly understood contexts. Through personal-

ized presentation of recommender messag-

es the system utilizes the most appropriate

format and therefore influences behavior.

The focus of our next stage of research is

to further our understanding of this con-

cept. This will enable a revision of the log-

ic within our basic trait model with the uti-

lization of additional context introducing a

greater certainty when considering options

for engagement.

6 CONCLUSION

In this paper we have addressed the fol-

lowing. 1) Whether an understanding of

user behavior traits and user affect can be

used to determine how best to present in-

formation in order to increase the likeli-

hood of m-commerce engagement. 2)

Whether context-aware recommender sys-

tems can adopt behavior trait models

alongside specific user context to deter-

mine how and when to present recommen-

dations.

For our first research question we conclude

that if the affective state of the user is

known then behavior traits can be used to

inform how best to present information to

a mobile user. We have shown that mes-

sages presented ‘in the wild’ via a smart-

phone which employ methods of mental

imagery, low effort, low risk and positive

appeals produce increasing levels of posi-

tive perception as user mood improves.

Though our results suggest that the same

does not apply to detailed analytical pro-

cessing, higher effort, fear appeals and

higher risk methods they do still provide

support for a basic behavior trait model.

We conclude that context-aware recom-

mender systems can adopt behavior trait

models to determine how to present a rec-

ommended item. We have also indicated

that with a greater understanding of the

impact of other user contexts then detailed

IF affective_state = positive

THEN using {

mental imagery

positive appeal

low risk

low effort }

ELSE

THEN consider_using {

detail processing

fear appeal

higher risk

higher effort }

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 70

Page 13: Message Perception within Context-Aware Recommender Systems

analytical processing, higher effort, fear

appeals and higher risk methods could also

be used within message customisation and

the presentation of recommender items.

Therefore our closing conclusion is that a

context-aware recommender system armed

with a comprehensive behavior trait model

will be able to determine the how and when

to present recommendations for optimized

engagement and thus improve return on

advertisement investment in m-commerce

contexts.

Future work will be directed towards strat-

egies of engagement and message person-

alization within recommender systems.

Our focus will be within context-

awareness, and the understanding of user

behavior within contrasting situations.

REFERENCES

[1] G. Adomavicius and A. Tuzhilin, “Context-

aware recommender systems,” in

Recommender Systems Handbook, Springer

US, 2011, pp. 217–253.

[2] N. D. Lane, E. Miluzzo, H. Lu, D. Peebles,

T. Choudhury, and A. T. Campbell, “A

Survey of Mobile Phone Sensing,”

Commun. Mag. IEEE, vol. 48, no. 9, pp.

140–150, 2010.

[3] B. A. Martin, “The Influence of Gender on

Mood Effects in Advertising,” Psychol.

Mark., vol. 20, no. 3, pp. 249–273, Mar. 2003.

[4] A. J. Rohm and V. Swaminathan, “A

Typology of Online Shoppers Based on

Shopping Motivations,” J. Bus. Res., vol.

57, no. 7, pp. 748–757, Jul. 2004.

[5] A. Smith and L. Sparks, “‘It’s Nice to Get a

Wee Treat if You've Had a Bad Week’:

Consumer Motivations in Retail Loyalty

Scheme Points Redemption,” J. Bus. Res.,

vol. 62, no. 5, pp. 542–547, May 2009.

[6] S. Youn and R. J. Faber, “Impulse Buying -

its Relation to Personality Traits and Cues,”

in Advances in Consumer Research, 2000, vol. 27, pp. 179–185.

[7] J. M. Otero-López and E. Villardefrancos

Pol, “Compulsive Buying and the Five

Factor Model of Personality: A Facet

Analysis,” Pers. Individ. Dif., vol. 55, no. 5, pp. 585–590, Sep. 2013.

[8] J. Myers and S. Sar, “The Influence of

Consumer Mood State as a Contextual

Factor on Imagery-Inducing

Advertisements and Brand Attitude,” J.

Mark. Commun., no. 3 (ahead of print), pp.

1–16, Feb. 2013.

[9] S. Brave and C. Nass, “Emotion in Human

– Computer Interaction,” in The Human-

Computer Interaction Handbook:

Fundamentals, Evolving Technologies and

Emerging Applications, no. Cmc, 2002, pp. 81–96.

[10] H. Lu, A. T. Campbell, and D. Gatica-

perez, “StressSense: Detecting Stress in

Unconstrained Acoustic Environments

using Smartphones,” in Proceedings of the

2012 ACM Conference on Ubiquitous

Computing, 2012, pp. 351–360.

[11] K. Yang and H.-Y. Kim, “Mobile Shopping

Motivation: an Application of Multiple

Discriminant Analysis,” Int. J. Retail

Distrib. Manag., vol. 40, no. 10, pp. 778–

789, 2012.

[12] A. Bhatnagar, S. Misra, and H. R. Rao, “On

Risk, Convenience, and Internet Shopping

Behavior,” Commun. ACM, vol. 43, no. 11, pp. 98–105, Nov. 2000.

[13] A. Holmes, A. Byrne, and J. Rowley,

“Mobile Shopping Behavior: Insights into

Attitudes, Shopping Process Involvement

and Location,” Int. J. Retail Distrib. Manag., vol. 42, no. 1, pp. 25–39, 2014.

[14] J. E. Burroughs, “Product Symbolism, Self

Meaning, and Holistic Matching: the Role

of Information Processing in Impulsive

Buying,” Adv. Consum. Res., vol. 23, no.

eds. Kim P. Corfman and John G. Lynch

Jr., Provo, UT: Association for Consumer Research, pp. 463–469, 1996.

[15] R. . Suinn, Psychological Techniques for

Individual Performance. New York: Macmillan Press, 1990, pp. 492–506.

[16] G. Y. Lee, “The Effect of Shopping

Emotions and Perceived Risk on Impulsive

Buying: The Moderating Role of Buying

Impulsiveness Trait,” Seoul J. Bus., vol. 14, no. 2, pp. 67–92, 2008.

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 71

Page 14: Message Perception within Context-Aware Recommender Systems

[17] G. Hastings, M. Stead, and J. Webb, “Fear

Appeals in Social Marketing: Strategic and

Ethical Reasons for Concern,” Psychol.

Mark., vol. 21, no. 11, pp. 961–986, Nov. 2004.

[18] M. R. Robberson and R. W. Rogers,

“Beyond Fear Appeals: Negative and

Positive Persuasive Appeals to Health and

Self-Esteem,” J. Appl. Soc. Psychol., vol. 18, no. 3, pp. 277–287, 2006.

[19] J. E. Escalas, “Imagine Yourself in the

Product: Mental Simulation, Narrative

Transportation, and Persuasion,” J. Advert.,

vol. 33, no. 2, pp. 37–48, 2004.

[20] K. S. . Yuen and T. M. . Lee, “Could mood

state affect risk-taking decisions?,” J.

Affect. Disord., vol. 75, no. 1, pp. 11–18, Jun. 2003.

[21] A. M. Isen, Positive Affect and Decision

Making. In M. Lewis & J. Haviland-Jones

(Eds.), Handbook of emotions, 2nd ed.,.

New York: Guilford, 2000, pp. 417–435.

[22] V. Augusta, K. Church, E. Hoggan, and N.

Oliver, “A Study of Mobile Mood

Awareness and Communication through

MobiMood,” in Proceedings of the 6th

Nordic Conference on Human-Computer

Interaction: Extending Boundaries, 2010,

pp. 128–137.

[23] M. P. Gardner, “Mood States and

Consumer Behavior: A Critical Review,” J.

Consum. Res., pp. 281–300, 1985.

[24] D. Lottridge, M. Chignell, and A. Jovicic,

“Affective Interaction: Understanding,

Evaluating, and Designing for Human

Emotion,” Rev. Hum. Factors Ergon., vol.

7, no. 1, pp. 197–217, Aug. 2011.

[25] D. T. Wegener, R. E. Petty, and D. J. Klein,

“Effects of Mood on High Elaboration

Attitude Change: The Mediating Role of

Likelihood Judgments,” Eur. J. Soc.

Psychol., vol. 24, no. 1, pp. 25–43, Jan. 1994.

[26] C.-L. “Eunice” Liu, R. R. Sinkovics, N.

Pezderka, and P. Haghirian, “Determinants

of Consumer Perceptions toward Mobile

Advertising - A Comparison between Japan

and Austria,” J. Interact. Mark., vol. 26, no.

1, pp. 21–32, Feb. 2012.

[27] R. J. Lutz, S. B. Mackenzle, and A. Ag,

“Attitude Towards the Ad as a Mediator of

Advertising Effectiveness: Determinants

and Consequences,” Adv. Consum. Res.,

vol. 10, no. 1, pp. 532–539, 1983.

[28] Z. Hadija, S. B. Barnes, and N. Hair, “Why

We Ignore Social Networking

Advertising,” Qual. Mark. Res. An Int. J., vol. 15, no. 1, pp. 19–32, 2012.

[29] A. Kleinsmith and N. Bianchi-Berthouze,

“Affective Body Expression Perception and

Recognition: A Survey,” IEEE Trans.

Affect. Comput., pp. 1–1, 2012.

[30] A. Bertron, M. Petry, R. Bruner, M.

Mcmanis, D. Zabaldo, S. Martinet, S.

Cuthbert, D. Ray, K. Koller, M.

Kolchakian, and S. Hayden, “International

Affective Picture System ( IAPS ):

Technical Manual and Affective Ratings

Lang , P . J ., Bradley , M . M ., & Cuthbert

, B . N . NIMH Center for the Study of

Emotion and Attention 1997 with the

assistance over the years of . - Mark

Greenwal,” 1997.

[31] R. J. Lavidge and G. a. Steiner, “A Model

for Predictive Measurements of Advertising

Effectiveness,” J. Mark., vol. 25, no. 6, p. 59, Oct. 1961.

[32] S. B. Mackenzie, R. J. Lutz, and G. E.

Belch, “The Role of Attitude toward the Ad

as a Mediator of Advertising Effectiveness:

A Test of Competing Explanations,” J.

Mark. Res., vol. 23, no. 2, pp. 130–143,

1986.

[33] The Nielsen Company, “Global Trust in Advertising and Brand Messages,” 2013.

[34] J. Park, S. J. Lennon, and L. Stoel, “On-line

Product Presentation: Effects on Mood,

Perceived risk, and Purchase Intention,”

Psychol. Mark., vol. 22, no. 9, pp. 695–719, Sep. 2005.

[35] T. Park, R. Shenoy, and G. Salvendy,

“Effective Advertising on Mobile Phones: a

Literature Review and Presentation of

Results from 53 Case Studies,” Behav. Inf.

Technol., vol. 27, no. 5, pp. 355–373, Sep.

2008.

[36] R. Likamwa, Y. Liu, N. D. Lane, and L.

Zhong, “Can Your Smartphone Infer Your

Mood?,” in PhoneSense workshop, 2011, pp. 1–5.

[37] A. Mehrabian, “Pleasure-Arousal-

Dominance: A General Framework for

Describing and Measuring Individual

Differences in Temperament,” Curr.

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 72

Page 15: Message Perception within Context-Aware Recommender Systems

Psychol., vol. 14, no. 4, pp. 261–292, Dec.

1996.

[38] S. Marsella, J. Gratch, P. Petta, and E. A.

Eds, “Computational Models of Emotion,”

in A Blueprint for Affective Computing - A Sourcebook and Manual, 2010, pp. 21–46.

[39] M. Bradley and P. J. Lang, “Measuring

Emotion: The Self-Assessment Manikin

and the Semantic Differential,” J. Behav.

Ther. Exp. Psychiatrysychiatry, vol. 25, no. 1, pp. 49–59, 1994.

[40] P. L. Wright, “The Cognitive Processes

Mediating Acceptance of Advertising,” in

Journal of Marketing Research, 1973, pp.

53–62.

[41] M. M. Tsang, S. Ho, and T. Liang,

“Consumer Attitudes Toward Mobile

Advertising: An Empirical Study,” vol. 8, no. 3, pp. 65–78, 2004.

[42] S. Rodgers and E. Thorson, “The

Interactive Advertising Model: How Users

Perceive and Process Online Ads,” J.

Interact. Advert., vol. 1, no. 1, pp. 42–61, 2010.

[43] K. Fiedler, S. Nickel, J. Asbeck, and U.

Pagel, “Mood and the Generation Effect,”

Cogn. Emot., vol. 17, no. 4, pp. 585–608,

Jan. 2003.

[44] M. Lewis, J. M. Haviland-jones, L. F.

Barrett, S. Rick, and G. Loewenstein, “The

Role of Emotion in Economic Behavior,” in Handbook of Emotions, 2008, pp. 149–150.

[45] T. Kujala, Y. Shtyrov, I. Winkler, M.

Saher, M. Tervaniemi, M. Sallinen, K.

Reinikainen, and R. Näätänen, “Long-term

Exposure to Noise Impairs Cortical Sound

Processing and Attention Control,”

Psychophysiology, vol. 46, no. 6, pp. 875–881, 2004.

[46] S. P. Banbury and D. C. Berry, “Office

Noise and Employee Concentration:

Identifying Causes of Disruption and

Potential Improvements,” Ergonomics, vol. 48, no. 1, pp. 25–37, Jan. 2005.

[47] F. Lederbogen, P. Kirsch, L. Haddad, F.

Streit, H. Tost, P. Schuch, S. Wüst, J. C.

Pruessner, M. Rietschel, M. Deuschle, and

A. Meyer-Lindenberg, “City Living and

Urban Upbringing Affect Neural Social

Stress Processing in Humans,” Nature, vol. 474, no. 7352, pp. 498–501, Jun. 2011.

[48] S.-H. Jeong and Y. Hwang, “Does

Multitasking Increase or Decrease

Persuasion? Effects of Multitasking on

Comprehension and Counterarguing,” J.

Commun., vol. 62, no. 4, pp. 571–587, Aug.

2012.

[49] M. De Sa and E. F. Churchill, “Mobile

Advertising: Evaluating the Effects of

Animation, User and Content Relevance,”

in Proceedings of the SIGCHI Conference

on Human Factors in Computing Systems, 2013, pp. 2487–2496.

Proceedings of the Third International Conference on E-Technologies and Business on the Web, Paris, France 2015

ISBN: 978-1-941968-08-6 ©2015 SDIWC 73