Phone-based Data Collection for Consumer Behavior Research

parchedmoosupElectronics - Devices

Nov 29, 2013 (3 years and 6 months ago)

107 views

IEEE TRANSACTIONS ON

JOURNAL NAME, MANUS
CRIPT ID

1


Phone
-
based Data Collection for Consumer
Behavior Research

C
huang
-
W
en

You
, Hsin
-
Liu (Cindy) Kao
,
Bo
-
Jhang Ho,
Y
u
-
H
an

Chen
, Wei
-
Fehng Wang, Hao
-
Hua
Chu, Lien
-
Ti Bei, Ming
-
Syan Chen

Abstract

Systematically and quantitatively determining patterns in consumer
flow

is an important question in consumer
behavior research. Identifying these patterns can facilitate understanding of where

and

when consumers purchase products and
services at
p
hysical

retail shops. Collecting naturalistic data on real consumers who shop at retail stores is one of the most
challenging and expensive

aspects of
consumer behavior studies. This paper introduces a phone
-
based data collection method
for consumer behavi
or research. The proposed method specifically targets local residents shopping at neighborhood
convenience stores. This study develops a phone
-
based data collection system
,
called ConvenienceProbe
,

and deploys and
tests the system by collecting real custom
er flow data in neighborhood convenience stores. Results show that the consumer
flow data collected from the ConvenienceProbe system is comparable to that from a traditional face
-
to
-
face interview method.

Index Terms

Phone

sensing, data collection

and anal
ysis,
consumer behavior
and marketing
research
.

——————————



——————————

1

I
NTRODUCTION
obile phones have become an indispensable part of
our everyday lives, as they go with us everywhere.
New mobile phones are equipped with sophistica
t-
ed sensing,

computing, and communication capabilities.
For example, a new smart phone
often

ha
s

a variety of
sensors, including GPS, accelerometer, digital compass,
Wi
-
Fi, and cell
-
ID sensors that can detect users’ locations
and movements. Due to the ubiquity of thes
e phones and
their sensory capabilities, it is possible to observe their
users
everywhere

and leverage and organize the resulting
data (from user
-
owned and

-
maintained phones) into
naturalistic and low
-
cost data collection systems that ca
p-
ture spatially re
levant information of human behavior.
This data collection of human behavior enables geograp
h-
ical and quantitative analysis of where, when, and how
people conduct their everyday activities invisibly and
non
-
intrusively, i.e., without disrupting human natur
al
behaviors.

This study develops a phone
-
based data collection a
p-
proach for
the

field of marketing research, and focuses on
determining spatial patterns in customer behaviors to
understand where, when, and how urban consumers visit
their neighborhood
C
on
v
enien
ce

S
tores (CVS). For exa
m-
ples, where do CVS customers come from? Where are the
gaps and overlaps in the market coverage of nearby and
competing CVS outlets? What is the cannibalization effect
of nearby CVS outlets competing for the same customer
base
in the same area?

1.1
Marketing Research and Retail Trade Area
Analysis

Since CVS outlets sell similar
merchandises
, marketing

and consumer behavior

research plays a particularly vital
role in determining their business success. Marketing r
e-
search is the
systematic approach of identifying, collec
t-
ing, and analyzing data relevant to marketing products
and services. Consumer research is a part of marketing
research that analyzes how changing marketing factors
affect customer behavior. This study provides a p
hone
-
based data collection system to gather data about custo
m-
ers’ CVS visits, including customer inbound/outbound
paths to CVS stores. Retail trade area analysis [
5
], which
is one kind of marketing research, provides a basis for
understanding, quantifying,

and visualizing customer
flow and movement in the area around a store. This i
n-
formation is critical for making business decisions such as
selecting the optimal store location, identifying compe
t-
ing stores, and placing outdoor advertisements. Since
consume
r flow and direction may change due to newly
opened or closed stores and changing composition of l
o-
cal residence, regular retail trade area analysis is nece
s-
sary to track a store’s current trade area and customer
flow.

xxxx
-
xxxx/0x/$xx.00 © 200x IEEE

————————————————



Chuang
-
Wen You
is with the
Research Center for Information Technology
Innovation
, Academia Sinica
. E
-
mail:
cwyou
@
citi.sinica.edu.tw
.



Hsin
-
Liu (Cindy) Kao

is with
the
Department of Computer Science and
Information Engineering
,
National Taiwan
Univ
.

E
-
mail:
b95701241
@
csie.ntu.edu.tw
.



Bo
-
Jhang Ho

is with the
Department of Computer Science and Information
Engineering
,
National Taiwan
Univ
.
E
-
mail:
b96118
@
csie.
ntu.edu.tw
.



Y
u
-
H
an

Chen

was

with the
Department of Computer Science and Info
r-
mation Engineering
,
National Taiwan
Univ
.

E
-
mail:
y
u-
han
@csie.ntu.edu.tw
.



Wei
-
Fehng Wang

is with
the
Department of Business Administration,
National ChengChi Univ
. E
-
mail:
92307036@nccu.edu.tw
.



Hao
-
Hua Chu

is with
the
Department of Computer Science and Info
r-
mation Engineering
, National Taiwan
Univ
. E
-
mail:
hchu
@
csie.ntu.edu.tw
.



Lien
-
Ti Bei is with the
Department of Business Administration, National
ChengChi Univ
. E
-
mail:
lienti@nccu.edu.tw
.



Ming
-
Syan Chen

is with the
Research Center for Information Technology
Innovation
, Academia Sinica
. E
-
mail:
mschen
@
citi.sinica.edu.tw
.


Manuscript received (insert date of submission if desired). Please note that all
acknowledgments
should be placed at the end of the paper, before the bibliography.

M

2

IEEE TRANSACTIONS ON

JOURNAL NAME, MANUS
CRIPT ID


1.2
Opportunities for Phone
-
based Dat
a Collection

Phones provide opportunities to outsource the process of
collecting customer flow data to local residents who own
and/or carry a mobile phone and are customers of neigh
-
borhood CVS stores. Outsourcing data collection to con
-
sumers can signific
antly reduce the cost of conducting
quantitative marketing studies. Phones provide oppo
r-
tuni
-
ties to automate the data collection process through
smart sensing, detecting, and logging of customers’ CVS
trips. Automating data collection makes it possible to

gather con
-
sumer behavior naturally without interrupting
users’ activi
-
ties, and reduces the under
-

or over
-
reporting and recall errors found in the traditional self
-
reporting, face
-
to
-
face interview, and surveying methods.

Traditional data collection me
thods for consumer b
e-
havior studies are labor
-
intensive [1]. In these approaches,
re
-
searchers recruit human field workers to observe cu
s-
tomers entering and leaving a store. Regular customers
are asked to fill out a questionnaire about their i
n-
bound/outbou
nd paths to the store and shopping beha
v-
ior. Another common human metering method involves
shadowing customers and observing their behavior over
time. Since human labor does not scale, data collection is
often an expensive part of consumer behavior researc
h.
This study proposes the ConvenienceProbe system to
reduce the cost of collecting consumer data. The proposed
system organizes phones as mobile sensors and gathers
customer flow data from phones.


The contributions of this study are two
-
fold:



This study

introduces a phone
-
based data collection
sys
-
tem that can significantly reduce the cost of data
collec
-
tion for consumer behavior researchers. This
system enables outsource and automation of co
n-
sumer data col
-
lection to everyday customers
thro
ugh automate
d sensing in their phones.



This study designs and implements the proposed
phone
-
based data collection system by collecting r
e-
al customer flow data to CVS outlets within an area
of interest. Re
-
sults show that the consumer flow
data collected from the phone
-
based method is
comparab
le to, and in some ways superior to, that
from a traditional face
-
to
-
face interview method.


The remainder of this paper is organized as follows.
The design and implementation of the ConvenienceProbe
system are described first. The next section compares data
collected from the ConvenienceProbe system and the tr
a-
di
-
tional face
-
to
-
face interview met
hod to identify real
cus
-
tomer flow data to CVS stores. The following section
dis
-
cusses the pros and cons associated with the Conve
n-
ience
-
Probe system and the traditional face
-
to
-
face inte
r-
view method. Finally, this study reviews related work
and draws co
nclusions and directions for future work.

2

S
YSTEM

A
RCHITECTURE

AND

D
ESIGN

The ConvenienceProbe system is based on the client
-
server architecture shown in Fig. 1. Phones carried by
participants run a client
-
side phone application that sen
s-
es participants’
consumer behavior. Phones periodically
transmit consumer behavior data to a backend server
through the phones’ wireless networks.

The ConvenienceProbe

system operates according to the
following steps. (1) Participants downloaded the phone
application to their phones and ran it in the background.
Participants carried their phones while performing no
r-
mal daily routines, including trips to neighborhood CVS

outlets. (2) The phone application implements automated
sensing, analyzing data from the phone’s location &
movement sensors and to detect visits to CVS outlets. The
phone application also logs CVS trip data in the phone.
(3) Participants periodically upl
oad CVS trip data to a
data repository on a backend server. To motivate partic
i-
pants to upload their consumer data, consumer behavior
researchers can set reward policies that determine how
much money participants earn given the quality and
quantity of data

uploaded from their phones to the ser
v-
er. (4) Optionally, participants can earn extra money by
correcting any mistakes made by the phone’s automated
sensing. (5) Consumer behavior researchers can view data
collection progress made by each and/or all parti
c
i-
pant(s), and then optionally consider dropping any pa
r-
ticipant not contributing data. (6) The server implements
Backend Server

Set incentive
policy

Participate in
data collection

Mobile phones

Phone application

Everyday consumers

Web portal

Data

Repository

Micropayments

Upload consumer
behavior data

Consumer researchers

Obtain consumer
data

Location & mov
e-
ment sensors

Fig. 1.
Client
-
server architecture of the ConvenienceProbe system.


Phone’s location and movement sensors

Phone interface

CVS trip detection

Optimal sensor selection for battery eff
i-
ciency

Wi
-
Fi sensor

GPS sensor

Accelerometer

Cell
-
ID sensor

Fig.
2
.
Design of the phone application
.


AUTHOR ET AL.: TITL
E

3


data visualization tools, allowing consumer behavior r
e-
searchers to view spatial patterns in customer flow beha
v-
iors from data stored in the s
erver’s data repository. The
following subsections explain the details of the phone
application and the backend server.

2.
1

Phone Application

Figure 2 shows the design of the phone application
that

includes the following two modules:
(1)
phone
user interfa
ce and (2)
automated CVS trip detection with
battery
-
efficient sensor selection
.

Participants us
e

the
phone user interface

to upload any collected data to a
server and/or
provide

meta
-
data description
s of

their
CVS visits. The mobile phone implements
automated

CVS trip detection

and logs any spatial data associated
with detected CVS trips.
D
ata collection software pe
r-
forms
optimal sensor selection

to conserve the phone

s
battery and
reduce
energy consumption when a partic
i-
pant is not within the target
area and/or not visiting any
neighborhood CVS stores. For additional
energy

saving
s
,
the

mobile phone
s

offload processing of i
n-
bound/outbound customer path reconstruction to
a

back
-
end server.
The proposed

phone application
currently
supports
HTC Andr
oid s
martphone series,
including

HTC Magic, Hero, Legend, and Desire

models
.
The fo
l-
lowing subsections explain each phone application mo
d-
ule in detail.


2.1.1 Phone user interface

A

simple

phone
user interface
that

allow
s

participants to
perform data uploading
and insert meta
-
data
descriptions
with

only

a few
button
clicks
. The user interface
serves
two main functions:
(1)
take a
photograph
of
the

CVS
purchase
receipt
as evidence

of a

store visit
,

and (2)
u
p-
load
CVS trip
data

logged in a phone to a server

as shown
in Fig. 3 (a)
.
To correct any mistak
es

made by the phone

s
automated CVS trip detection, p
articipants

match
ed
each
receipt photograph to a list of

phone
-
detected CVS visits

(Fig. 3 (b))
.
T
he automated detection algorithm
counts
a
ny receipt photog
raph without a matching phone
-
detected CVS visit (i.e., a false negative in which the
phone did not detect a CVS visit but in fact there was a
visit)
or

any phone
-
detected CVS visit without a matching
receipt photograph (i.e., a false positive in which the

phone detected a CVS visit but in fact there was no visit)
as an error. The user interface

provides

two
uploading
methods

to
transfer

CVS visit
data

from a phone to the
server
. The first

uploading method uses a direct wireless
Internet connection on a pho
ne to transfer data to the
server. In the absence of a direct wireless Internet conne
c-
tion on a phone, the second uploading method uses a
USB connection to a PC, in which the PC first retrieves
data from the phone and then transfers data to the server.

2.1
.2
Automated CVS trip detection with battery
-
efficient sensor selection

Running data collection software consumes a phone

s
limited battery.
Since

participants
must

use their phones


battery for normal day
-
to
-
day communications, battery
management
is
a key issue on phone
s
. In particular, pa
r-
ticipants
may

complain

and/or refuse to participate in the
data collection process if the data collection software co
n-
sumed
a lot of

battery
power,
but not collect
enough
us
e-
ful data
to earn

rewards. In other words, the data colle
c-
tion

software
must

scale down

its battery consumption
during times of zero/low probability of collecting useful
data.

This study presents a

battery
-
efficient sensor selection
algorithm
t
o achieve this scaling
-
down goal by
conside
r-
ing

sensors


accuracy
-
energy
tradeoff
, where a
ccuracy
represents a sensor

s positional accuracy
,

and energy re
p-
resents the sensor

s energy consumption. For example,
the GPS sensor gives accurate location
information,
but
consumes
a lot

energy, whereas the cell
-
ID sensor pr
o-
vides
a rou
gh

location but
uses

little energy. When a
phone is outside the area of interest, the probability of
collecting useful data is almost zero. Hence, the sensor
selection algorithm turns on only the low
-
power cell
-
ID
sensor
, which

outputs coarse location sufficient to dete
r-
mine when a participant enters the area of interest.
The
sensor algorithm simultaneously

turns off all other se
n-
sors to conserve battery

power
.

Fig 3. Phone’s main and labeling interfaces.

The main interface (Fig. 3(a)) shows a photo receipt capturing button, a data uploading
button, and an exit button. The labeling interface (Fig. 3(b)) shows a list of phone
-
detected store visits from which a participant
matches each receipt photo to.


(a) The main interface

(b) The labeling interface


4

IEEE TRANSACTIONS ON

JOURNAL NAME, MANUS
CRIPT ID


The battery
-
efficient sensor selection algorithm consi
d-
ers the
foll
owing

location and movement sensors co
m-
monly found in smart phone
s
: cell
-
ID, accelerometer,
Wi
-
Fi
, and GPS sensors, which are listed in order of increa
s-
ing energy consumption.
The

CVS trip detection alg
o-
rithm uses each of these sensors
in the following way
s
. (1)
The cell
-
ID sensor detects a phone entering the area of
interest by matching the phone
-
detected cell
-
ID(s) to
those belonging to the area. (2) The accelerometer sensor
detects the presence/absence of user movement.
The
p
resence of user movement trig
gers
continuous

location
update,
while its

absence
pauses

location update
s
. (3) The
Wi
-
Fi

sensor
[
4
]

identifies
when
a participant
transitions

between places (e.g.,
moves

from an office to a CVS ou
t-
let)
or stops in one

place (e.g., shopping at a CVS outlet
).

The p
hone
-
received
Wi
-
Fi

signature is also used to reco
g-
nize a specific CVS outlet by matching the phone
-
received
Wi
-
Fi

signature to the
Wi
-
Fi

APs previously profiled and
located nearby the CVS outlet. (4) The GPS sensor
ident
i-
fies
inbound/outbound path
s to a CVS store. GPS signals
from a phone can also be used to detect indoor/outdoor
transition
s
.

W
hen a participant
/phone

enters a building,
the GPS signals become weak or undetectable.

Figure

4

illustrates

the battery
-
efficient sensor selection
algorithm
,

which chooses the optimal sensor(s) according
to the phone

s spatial context. The three
algorithm

states
are defined as follows.

(
S1) Outside the area of interest
:

T
he

sensor selection

alg
o-
rithm
turns on

only the low
-
power cell
-
ID

sensor
,

which
checks

if
the phone

enters the area
of interest

based on the

cell
-
ID
(s)
in the

area
.

(S2) Outdoor and inside the area of interest
:

The sensor
selection algorithm turns
on
all sensors except
the
acce
l-
erometer. These selected sensors are used as follows. The
cell
-
ID sensor checks

when

the phone

leaves

the area
of
interest

based on

the
dis
appearance of
local
cell
-
ID
(s)
.
The
GPS sensor de
tects any transition from outdoor
s

to indoor
s
.

The
GPS sensor also logs any inbound/outbound path to a
place such as a CVS store.
Wi
-
Fi

sensor identifies a specific
CVS store by matching the phone
-
received
Wi
-
Fi

signatures
to those belon
ging to

a CVS outle
t.

(S3) Indoor and inside the area of interest
:

The

sensor sele
c-
tion

algorithm

turns on the cell
-
ID, accelerometer, and
GPS sensors. Cell
-
ID sensor
determines
when

the phone
leaves

the area
of interest

based on

the
dis
appearance of
local
cell
-
ID
(s)
. The
accelerometer continue
s

or pauses

location update
s

based on

the
presence/absence of user
movement
, respectively.
In the presence of user mov
e-
ment, the algorithm turns on the GPS sensor at a low
sampling rate, e.g., every other minute, to detect any
transit
ion from
indoors

to
the outdoors
.


T
ABLE

1

C
ONFUSION MATRIX MEAS
UR
ES

THE ACCURACY OF
CVS

IN
-
STORE VISIT
DETECTION


Detection

Actual

Visits

Non
-
visits

Visits

3
45

23

Non
-
visits

24

218


This study includes an

experiment to determine the a
c-
curacy of
the
proposed

battery
-
efficient CVS trip dete
c-
tion algorithm. Data
was obtained from
42 participants

and

368 CVS visit logs
in

the user study (described in the
next section)
.
Since a CVS visit leaves GPS log trace
around locations of CVS outlets,

this study

define
s

dete
c-
tion accuracy as how well the CVS trip detection alg
o-
rithm correctly classified events
in which

participants
only passed by a CVS store without entering it (defined as
a non
-
visit) from events where participants entered a CVS
store (defined a
s a visit).
Participants
were

asked to take
photograph
s

of
purchase receipt
s

as ground truth ev
i-
dence of

CVS visit
s
.
Any

unmatched receipt photograph

or

any
unmatched phone
-
detected visit

constitutes

an
error

in the CVS visit
detection algorithm.

Table
1

shows the accuracy result in a confusion matrix.
The detection accuracy of CVS visits was 9
3
.
75
%, which
was better than the detection accuracy of CVS non
-
visits
at 9
0.08
%. Most of the detection errors came from partic
i-
pants standing near the entrance to a
CVS store where the
phone
-
received Wi
-
Fi signature is similar to the in
-
store’s
Wi
-
Fi signature
,

thus causing detection errors. A 90% plus
CVS detection accuracy rate is sufficient for consumer
behavior research.

In a
ctual user experiment
s

with 42
particip
ants,
no

participant
s

reported
that
a
phone
fully
charged in the morning ran out of battery at the end of
the

day using
the proposed

battery
-
efficient sensor sele
c-
Indoor to out
door
transition

(
S
1) Outside the area of interest

[On] Cell
-
ID sensor

[Off] Accelerometer

[Off] WiFi sensor

[Off] GPS sensor

(
S
2) Outdoor and inside the area of interest

[On] Cell
-
ID sensor

[Off]
Accelerometer

[On] WiFi sensor

[On] GPS sensor

(
S
3) Indoor and inside the area of interest

[On]
Cell
-
ID sensor

[
On@low
_rate] Accelerometer

[Off] WiFi sensor

[On@
low
_rate] GPS sensor

Upload CVS visit data

M
ove inside the area of interest


M
ove outside the
area of interest

M
ove outside
the area of
interest


Outdoor to in
door
transition

Fig
.

4
. The state diagram of the battery
-
efficient sensor sele
c-
tion

algorithm
. Three states are S1(outside the area of interest),
S2(inside the area of interest and outdoor) and S3(inside the
area

of interest and indoor). Each state shows the selected
sensor(s) as [
O
n] and unselected sensor(s) as [
O
ff].

AUTHOR ET AL.: TITL
E

5


tion algorithm.

2.1.3
Data Uploading

To protect
the
location privacy of participants,
the pho
ne
only
upload
s

the CVS trip data to the server. From the
phone user interface, participants click
ed

the uploading
button to initiate any data uploading
. This

process r
e-
quired participants to
explicit permission to the phone
application
for

each

data uploading operation.

2.
2

Back
-
end Server

Figure
5

shows the
design

of t
he backend server
, which

includes (1) the web portal, (2) data repository, and (3)
data analysis and visualization tools. The following su
b-
sections explain each server component i
n detail.

2.2.1 Web portal

Participants could
browse the portal’s web pages
for

i
n-
formation about
the

consumer behavior research and its
data collection process, including
a
description

of
the
purpose of this study, user qualifications to participate in
this study, incentive policies, etc. After individuals agre
ed

to participate in this study, the web portal ask
ed

them to
complete questionnaires to
determine

their qualificatio
n
for
involvement
. The questionnaire

used in this study
asked
whether they live and work within the area of i
n-
terest, how frequent
ly

they visit neighborhood CVS ou
t-
lets, whether they own and carry smart phones capable of
running the ConvenienceProbe phone
application, etc.
Participants that met the necessary qualifications were
asked
to fill out their contact and user profile information
and to create individual accounts
from which

they
could

upload data and receive micropayments. Finally, the web
portal as
k
ed

them to sign a consent form agreeing to r
e-
lease their data. Each time a participant upload
ed

data to
the server, the website
calculated

the amount of money
earned and the amount of data contributed to the server’s
data repository.

2.2.2 Data repository

The data repository provides a centralized storage of all
CVS visit data uploaded from participants. The CVS trip
data uploaded
in this study includes

(1) raw sensor data
including timestamp, GPS coordinates, cell
-
ID log, and
received Wi
-
Fi signatures, an
d (2) phone
-
detected store
visits and any meta
-
data description
s

provided

by parti
c-
ipants.
M
arketing research and retail trade area analysis
retriev
ed

data from the data repository
for

customer data
analysis and visualization.

Customer path reconstruction.

For each phone
-
detect
CVS trip, the system reconstructs
the corresponding co
n-
sumer path
. Each consumer path has three elements. (1)
An inbound path is the store
-
arrival path taken by a cu
s-
tomer starting from a previous destination place (e.g.,
participant
’s home or office building)
until

arriving at the
store. (2) In
-
store time is the amount of time that a cu
s-
tomer stays at the store. (3) An outbound path is the
store
-
departure path taken by a customer leaving the
store
for

his/her next destination. GPS da
ta from phones
was used to plot

these consumer paths and form routes.

Retail trade area visualization.

Figure
8
(a) shows an e
x-
ample of a retail trade area map computed from the
Bounding Wedge
-
Casting method [
9
]. This method d
i-
vides a store’s surrounding a
rea into directional wedge
sectors. The store is the hub at the center of all wedge se
c-
tors. For example, the retail trade area map in
Fig.

8
(a)
contains twelve 30
-
degree wedge sectors. Each wedge
sector grows from the store location in the center ou
t-
wards

to cover locations of additional customers until the
cumulative number of customers exceeds a threshold
,

such as 80% of customers.

3

U
SER
S
TUDY

The effectiveness of the ConvenienceProbe system was
tested in a user study that collected real customer flow
d
ata from three competing CVS stores situated within the
same area. To compare the ConvenienceProbe system
with the traditional data collection method, a pen
-
and
-
paper survey was also conducted with the customers of
these CVS stores. This section describes
the design and
experimental results of this comparative user study,
which were gui
ded by the following inquiries:




What was the relative data quality and quantity
collected from the ConvenienceProbe system co
m-
pared to the traditional face
-
to
-
face
interview?



What was the trade area analysis quality obtained
from the ConvenienceProbe system compared to the
traditional face
-
to
-
face interview?


Fig.
2
.
Design of the phone application
.


Trade area visualization tools

Consumer path reconstruction

Data repository

(
T
imestamp, GPS coordinates, Cell
-
ID
logs,Wi
-
Fi signatures)

Web Portal

Participating
Phones

Consumer behavior
researchers


Trade area maps


Fig 5.
Design

of the back
-
end server
.

6

IEEE TRANSACTIONS ON

JOURNAL NAME, MANUS
CRIPT ID


Exactly how well the ConvenienceProbe

system collects
data was evaluated by comparing its spatial patterns with
those obtained from the traditional face
-
to
-
face interview.
To make a meaningful comparison, this comparison
study emphasizes the aspect of data collection methods,
i.e., the tradit
ional face
-
to
-
face interview method (called
the inter
-
view method from here) vs. the Convenie
n-
ceProbe method (called the phone method from here),
while maintaining consistency in other aspects, such as
the same CVS outlets in the focal neighborhood, the sa
me
qualification criteria for selecting participants, and the
same data analysis me
-
thod on the customer flow (i
n-
bound/outbound paths
and source/destination points).

3.1 Physical Settings

The study area was a

highly
-
populated
triangular
city
block near
a
u
niversity campus
. Figure
6

marks this study
area in
brown
boundary lines.
T
his triangular area mea
s-
ure
d

approximately

0.128 square kilometer, and the tr
i-
angular edges
were

approximately 470, 540, and 640 m
e-
ters

long
. This focal area
contained

three competing CVS
outlets
,

whose locations are marked on Fig.
6

as
stores
(A),
(B)
, and
(C).

3.2 Participants

T
he
phone
campaign

recruited
42

participants

who either
already owned HTC Android phones

capable of running
our phone application o
r
borrowed
a

compatible HTC
Android
phone

from us for the duration of
the

study
.
The

age distribution

of participants ranged be
tween 18 and 53
,
and
the
average

age
was 25
. Their occupations included
students, clerks
, sales, engineers
,
housekeepers,
etc
. All
participants were residents

of, or worked in,
the
focal

area
,
and

regularly patronized the selected
CVS
stores.
Co
m-
pensation to participants was

a fixed
weekly
payment

with a guaranteed base
NT$
500

(
about

US$17)

plus an

extra daily bonus
NT$
30
(
about

US$1
)

directly
related
to
the number of days with
uploading CVS

visit

data
.

To
prevent
demand artifac
ts
,
the

maximum weekly
payment

was set at
NT$
700

(
about

US$
23.7
)

so that participants
would not increase their CVS visits due to the minor daily
incentive
.

The

4
2 participants

in this study recorded

3
68

visits to the three CVS stores

during 3 weeks
,
with

1
67
,
15
0
, and 5
1

visits to stores A, B, and C
,

respectively
.

The

interview campaign surveyed

90 frequent custo
m-
ers, or 30 customers per store, outside of the three CVS
stores.
T
o
reduce

the
effect

of the time
-
dependant factors
on
the profile of customers
, respondents were recruited
at

various hours of a day
(10am ~ 8
pm) and

both weekdays
and weekends
. The age

distr
ibution
of the

respondents

ranged between 12 and 73
, and the average

age
was 2
9
.
Their occupations
were similar to
the
occupations of
pa
r-
ticipants in the

ConvenienceProbe

campaign.

Table 2 lists the

participants in both phone and inte
r-
view
campaigns
. Two
-
tailed T
-
test results show n
o signi
f-
icant

differenc
es

between

these two

groups in
gender
(
p
=0.
619),
age

(
p
=0.
200
),

store visit frequency (
p
=
0.972
)
,
and
disposable income levels
1

(
p=
0.725
)
.

3.3
Procedures of the Phone Method

T
he procedure consisted of two
phases: (1) a screening
phase to select qualified subjects and collect background
information
,

and (2) the experiment phase. Consumers
were recruited via
an

Internet ad and snowball
sampling
.
The selection criteria were those who either lived or
worked in
the focal area,
and

frequently shopped at one
of the three CVS stores.
Q
ualified

participants

began pe
r-

1

According to the s
urvey of
f
amily
i
ncome and
e
xpenditure

conducted
by the
Directorate
General of Budget, Accounting and Statistics, Exec
u-
tive Yuan,

Taiwan, the monthly personal

disposable income (PDI) of 2009
in Taiwan is 642 USD.

T
ABLE
2

D
EMOGRAPHY OF PARTICI
PANTS IN
THE
C
ONVENIENCE
P
ROBE AND TRADITIONAL

HUMAN
-
INTERVIEW CAMPAIGNS
.

Data Collection
Methods

#
of
Subjects

Male
(Female)

Age

distr
i-
bution

# of
store
visits

Store visit
frequency (per
week)

D
isposable income levels v.s. the
number
of subjects

ConvenienceProbe
system

42

2
3 (1
9
)

18~53

(mean=25
,
std=
7
)

394


0.5~10
(
mean
=3.85,
std=2.26)

Less than Monthly PDI

More than monthly PDI

(34)

(8)


Traditional h
u-
man
-
interview

90

54 (36)

12~73

(mean=28,
std=1
3
)

90

1~20
(
mean
=3.55,
std=2.85)

Less than Monthly PDI

More than monthly PDI

(66)

(24)




Fig
.

6
.
The triangular dotted boundary lines mark the focal area
.
This area is adjacent to

a university campus

(X)
.
There are t
hree
CVS

stores:

(
A
) Family Mart, (
B
)

Xinsheng 7
-
11, and (
C
) Roos
e-
velt 7
-
11.
Other

landmarks near the
focal
area includ
e

G
ong
g
uan

(Y)

and Tai
-
P
ower

(Z)

metro rapid transit

(MRT)

stations.

AUTHOR ET AL.: TITL
E

7


forming data collection tasks and were asked to regularly
upload their
CVS visit
data.

In the screening phase, interested candidates first
co
m-
pleted
questionnaires on the web portal. The questions

(in
Table 3)

assessed

the candidates’ familiarity wi
th the dig
i-
tal devices
, i.e., whether they had basic technical skills,

and their da
ily consumption habits
, i.e.,

whether they
regularly patronized the selec
ted CVS stores. After they
passed the screening stage, candidates were asked to fill
out a pre
-
study questionnaire with their personal
info
r-
mation
. They also set up accounts for uploading data later.
Finally, q
ualified participants were asked to attend an
orientation

on how to use the
phone
-
based
system.

The experiment phase

ran for 2
-
3 weeks
, during which

participants carried phones that recorded i
n-
bound/outbound paths of their CVS trips
.
Participants

were
responsible

for
recharg
ing
the

phones a
t the end o
f
each day
.

3.4
The Interview Method

This procedure

involved two steps
.
(1)
Interviewers a
p-
proached c
ustomer
s

as

they left the

CVS

store
.
(2) If the
customers were

regular customer
s

of th
e

focal

CVS
store
,

interviewers
helped

customers
fill

out

a short questio
n-
naire about
the

origin
/
destination

points

of
their

current
store
visit and
other consumer

behavior.

3.5
Evaluation metrics and results for
inbound/outbound directions.

C
ustomer inbound/outbound directions
identified

which
rout
es

customers
t
ook

to enter

and
leave
each

store.

Fi
g-
ure
7

shows the location of stores A
and
B with 3
-
4 poss
i-
ble customer inbound/outbound directions entering

and
leaving these stores.

This
study defines

S
d

as a metric for measuring simila
r-
ity in customer inbound/outbound directions between
the
phone

and interview methods. The compositions and
calculations of this
S
d

v
alue are described as follows. (1)
A
ll
the CVS
store visits
were divided
into different flow
components
based on

customer

inbound/outbound d
i-
rections.
(2)
A

customer flow vector

for
each

CVS
store

was
generated
for
both

data collection
methods
. A customer
flow vector contains vector components whose scalar
values correspond to
the
percentage
s of customers ente
r-
ing/leaving the store from a specific direction
. Figure
7

shows that

each
store has two customer
flow vector
s from
the phone method (denoted as
𝑽






) and the interview
method (
𝑽




). Each customer flow vector
V

has vector
components corresponding to multiple flow directions:
𝑽









= (
D
1
, D
2
, .., D
m
).
For example,
s
tore
A

in
Fig.
7
(a) has
three

flow directions
(
D
1
, D
2
, D
3
)
. The scalar
value
D
1

i
s the percentage of customers entering/leaving
the store from/to the north direction. (3)
T
he Cosine
-
based correlation between the two customer flow vectors

(
𝑽







and
𝑽




)

was computed for each store
as
a

measure of similarity.



𝑆
𝑑
=
𝑽







𝑽




|
𝑽






|

|
𝑽




|



(1)


Table
4

shows the
similarity results
for

customer flow
vectors of
all

three CVS stores. Data were obtained by
analyzing customer inbound/outbound paths collected
from the phone method (Column 2) and the interview
method (Column 3). Column 4 calculates the Cosine
-
based correlation similarity scores (
S
d
) between the phone
and

interview methods
for

the inbound/outbound dire
c-
tions of three stores using Eq. (1). For all three stores, si
m-
ilarity scores are higher

than
92%.
The
similarity scores
f
or

store
s

A
and B

reach
near
ly

100%.

T
he average

sim
i-
larity score

among the three stor
es is
9
6%. For statistical
validation,
we conducted the C
hi
-
square
(
χ
2
)

equality of
proportion test

for equality of distributions on data o
b-
tained by the phone method (i.e., the observed) and the
interview method (i.e., the expected). Column 5 of Table
4

shows the
χ
2

and the corresponding
p

value for each store
(respectively,
χ
2
(
2
)=0.037,
p
=0.9822

for store A;
χ
2
(3)=0.020,
p
=0.9985 for store B;
χ
2
(2)=0.199,
p
=0.9053 for store C).
The
p

values are

all
close to 1 and
larger

than the conventio
n-
ally accepted s
ignificance level of 0.05
. So, the
null h
y-
pothesis that the two distributions are the same is

not

rejected

for each store. In other words, the C
hi
-
square
equality of proportion test

indicates a good fit between
the phone method (observed) and the interview

method
(expected) for all three stores.

(a)
Store A

(b)
Store
B

Fig
.

7
.
The inbound/outbound direction
s

of
store A and B
.
Store A
is

located at 3
-
way T
-
intersections

with three inbound/outbound
directions
, and

store B
is located in a
n alley

with four i
n-
bound/out
bound directions.

T
ABLE
3

Q
UESTIONS FOR SCREENI
NG SUBJECTS
.


ID

Questions

F
amiliarity
with the
digital d
e-
vices

C1

Do you usually use high
-
tech products?

C2

Do you use a smartphone as your pr
i-
mary phone?

C3

Which model is your smartphone? (HTC
Magic, Hero, Legend, Desire, iPhone or
others)

D
aily co
n-
sumption
habits

S1

Do you live or work in the
focal

area?

S2

Do

you patronize at least one target CVS
every three days?

S3

Please select the CVS

store
s that you
regularly patronize in your daily life?
(“none of them”, “Xinsheng 7
-
11”,
“Family Mart”, “Roosevelt 7
-
11” or
“more than one store”)



8

IEEE TRANSACTIONS ON

JOURNAL NAME, MANUS
CRIPT ID


3.6
Evaluation Metric and Results for Retail Trade
Areas

This study defines

three metrics to measure

the

similarity
of trade areas
obtained from the

phone and interview
methods. (1) Denote
overlapping ratio

as the
overlapped

trade area
between

the phone

and
interview methods
divided by the trade area of the interview method (as the
baseline area). (2) Denote
miss ratio

as the trade area
found in the interview method
,

but not in the phone
method
,

divided by the trade

area of the interview met
h-
od. (3) Denote

extra ratio

as the trade area
found

in the
phone method
,

but not in the interview method
,

divided
by the trade area of the phone method.

A h
igh overlapping percentage suggests that data co
l-
lected from the phone an
d interview methods have little
difference in term of analyzing trade area results. Recall
from
the
previous section

(Bounding Wedge
-
Casting
Method) that a retail trade area
includes

wedges
,

each
of
which

is a directional sector area that grows from the
st
ore’s location outwards to cover customers’ locations.
Figure
8

plots the conservative retail trade areas of stores
A,
and
B in the 30
-
degree wedge presentation.

Table
5

shows the similarity results indicated by

the

r
e-
tail trade areas of the three CVS stores. The overlapping
ratio

(Row 5)

is high at 8
6
.5%, suggesting that the phone
method was able to capture
more

trade areas
than

the
interview method. Similarly, the miss ratio (
Row

6) is low
at 1
3
.
5
%, suggesting that the phone method missed
fewer

trade areas
than

the interview method.

The extra ratio (
Row

7)
is

high at
77.8
%, suggesting that
the phone method captures a lot more trade area than the
interview method.
The

high extra ratio in Fig.
8

shows
that the trade areas of stores A and B obtained from the
phone method (upper two graphs in Fig.
8
) cove
r the
trade areas obtained from the interview method (middle

two

graphs in Fig.
8
).
The

lower
two

graphs in Fig.
8

show the

overlapping tr
ade areas of these two methods.
Note that the trade area graphs of store
C

are omitted due
to space limitation
s
.


The
reason
for the high extra ratio
was due
to
the

lim
i-
tat
ions of the interviewing method
.

Since a face
-
to
-
face
interview
is

limited to only a few minutes,
res
pondents
often used

intermediate
land
mark

points
on their
store
arrival
paths

as approximate location
s
of

their ac
tual
source points
. This

resulted in
mostly
smaller trade areas
from the interview method than those from the phone
method.
The

Discussion Section

presents

this limitat
ion in
greater detail
.

4

D
ISCUSSION

This section discusses
the
pros and cons associated with
the phone method (the ConvenienceProbe system) and
the interview method (the traditional face
-
to
-
face inte
r-
view).

4.1
Causes of Imprecise Data: Phone Sensory
Errors vs. Human Recall/Communication
Errors

T
he phone method collected more
elaborated

and
precise

consumer flow data than the interview method. For ma
r-
keting research, detailed and accurate customer
source

and
des
tination points and path
information

are critical
to

obtaining accurate trade area results and for understan
d-
ing the trade areas in the focal region.
There are

two re
a-
sons
for the different

data quality between these two
methods.

First, the interview meth
od relie
s

on human memory
and human spatial cognitive capabilities to recall and
communicate source
and

destination points during face
-
to
-
face interviews. Since face
-
to
-
face interview
s were

li
m-
ited to only a few minutes, respondents often used coll
o-
quially
-
labeled landmarks as their source/destination
locations, e.g., “nearby the McDonald restaurant
,
” “nea
r-
by the church
,
” “nearby the bus station
,
” etc., which were
verbally quick to communicate but lacked fine
-
grained
spatial accuracy. Additionally, responde
nts often had
T
ABLE

4

S
IMILARITY RESULTS
FOR

THE FLOW VECTORS
(
I
.
E
.,

A STORE

S
C
USTOMER
INBOUND
/
OUTBOUND DIRECTIONS
)

OF THE
PHONE

AND INTERVIEW METHOD
S FOR
ALL

THREE
CVS

STORES
.

S
IM
I-
LARITY SCORES
WERE

COMPUTED USING
E
Q
.

(1).


CVS
stores
Inbound

outbound
directions

Flow vector
from the
phone
method
(n=42, #vi
s-
its=3
68
)

Flow vector
from the
interview
method
(n=90,
#visits=90)

Similar
i-
ty score
(
S
d
)

χ
2
value

Store A



98.42%

0.037

D1

17.3%

13.4%


(p
=0.9822
)

D2

27.8%

36.6%



D3

54.7%

50.0%



Store B



98.5
6
%

0.020

D1

38.2%

32.2%


(p
=0.9985
)

D2

25.7%

25.5%



D3

17.9%

25.0%



D4

17.9%

17.8%



Store C



92.38%

0.199

D1

24.1%

24.3%


(p
=0.9053
)

D2

37.9%

55.1%



D3

37.9%

20.6%




T
ABLE

5

S
IMILARITY RESULTS BE
TWEEN THE TRADE AREA
S
(
I
.
E
.,

THE
AREA WHERE A STORE

S C
USTOMER
COME FROM
)

OF THE
PHONE

AND INTERVIEW METHOD
S FOR EACH OF THREE
CVS

STORES
.

S
IMILARITY IS MEASURE
D BY OVERLAPPING
,

MISS
,

AND EXTRA RATIOS
(
DEFINED IN THE PREVI
OUS PARAGRAPH
).



Store A

Store B

Store C

Phone method

trade
area (km
2
)
(n=42, visits=3
68
)

7.631

4.520

2.894

Interview method


trade

area (km2) (n=90, visits=90)

1.959

1.071

2.134

I
ntersection area (km
2
)

1.695

0.963

1.626

Overlapping

r
atio

(%)

86.5%

89.9%

76.2%

Miss
r
atio

(%)

13.5
%

10.1
%

23.8%

Extra ratio (%)

77
.
8
%

78.7
%

43.8%



AUTHOR ET AL.: TITL
E

9


difficulty communi
cating their store

arrival and
depa
r-
ture paths due to a lack of spatial orientation and dire
c-
tions.
E
ven with the help of a map, respondents had
diff
i-
culty

correctly locating themselves on the map and pi
n-
pointing travelled p
aths. In contrast, the phone method
recorded and mapped customer flow data using sensors
and software tools. Although
the

phones’ GPS sensors
in
this study had

an average positional error of 10
-
15 meters,
their sensory errors

were considerabl
y

less than im
precise
human communication.

Second, since face
-
to
-
face interviews took place
when

respondents
exited

the
stores, we could only ask r
e-
spondents
about

their
expected

destination

points

and

d
e-
parture paths, which might be different from
their

actual

destination points and departure paths.
The

phone met
h-
od did not have this problem
, as it tracked the
continuous
movement of consumers leaving stores.

4.2
Compensation Models: Payin
g Customers vs.
Paying Interviewers

The phone method place
s

the burden of data collection on
customers who owned, maintained, and operated the
data collection devices, i.e., their phones. In contrast, the
interview method place
s

the burden of data collecti
on on
interviewers who traveled to the stores and manually
collected data from the customers.
T
hese two methods
produced different compensation models
.
In the phone
method, most compensation went to paying the custo
m-
ers for their de
vice management

time and

effort
. In the
interview method, most compensation went to paying the
interviewers
for

their
time and
effort
.

Unlike

the interview method, the

phone method offers
additional flexibility and potential for extending its co
m-
pensation model to include
real
-
time and context
-
based
rewards such as in
-
store coupons, which could be deli
v-
ered over each participant’s phone and tailored specifica
l-
ly toward his/her location and historical consumer b
e-
havior.
T
his context
-
relevant coupon service could
i
n-
crease

the

incentive for participation in the study and
reduce the direct payouts needed
. This cost saving
could
be significant if the study involved a large number of pa
r-
ticipants. In other words, flexibility in the compensation
model in the phone method offers sca
lability potential
for

a large number of everyday users as
participants
.

5

R
ELATED
W
ORK

This section
review
s

traditional techniques
for

collecting
consumer behavior data
, and then
describ
es

related

stu
d-
ies apply
ing mobile and UbiComp

technologies in ma
r-
ket
ing and consumer research. Finally,
this section

di
s-
cuss
es

related phone
-
based behavior sensing systems.

5.1
Traditional Data Collection Techniques in
Marketing and Consumer Research

Traditional data collection techniques for consumer
flow
and trade area

studies
include human shadowing and
recordi
ng [
2
], or consumer surveys [
24
]. Human shado
w-
ing and recording techniques
require

human observers at
stores
,

who directly follow or indirectly survey ing
o-
ing/outgoing customers to determine their trajectories
up
on entering/leaving and/or their origins/destinations.
Human observers also track the number of inco
m-
ing/outgoing customers to determine the number of store
visits. Other human recording [
3
] or surveying techniques
record subtle consum
er

behavior
s,

but require intensive
human labors.

Other techniques for analyzing trade area are based on

the concepts of spatial monopoly
and

market penetration.
The concentric rings methods, drive time/distance pol
y-
gons or Thiessen (Voronoi) polygons, and prob
abilist
ic
trade area surfaces [
6
] are examples of the spatial mono
p-
oly approach. These techniques are often simplified and
(c) Store A (interview method)

(e
) Store A

(overlapping phone
/interview)


(a)
Store A

(phone method)

Store C

(b)
Store
B (phone method)

(d)
Store B (interview method)

(f
) Store
B (overlapping phone
/interview)


Fig
.

8
. The trade areas of
stores A

and B

obtained from

the
phone method (the upper graphs) or
interviewing

method (the
middle graphs)
.
The bottom row shows
overlapping

trade areas
of these two methods. The black regions indicate the intersection

regions between two methods.

10

IEEE TRANSACTIONS ON

JOURNAL NAME, MANUS
CRIPT ID


fail to incorporate the effects
of

competing
nearby stores.
The Huff model [
5
] is an example of the market penetr
a-
tion approach, which inclu
des the effect of competing
stores by modeling a spatial variation in the proportion of
households served by each surrounding store. However,
this technique requires a calibration phase in which
r
e-
sults of
customer survey

compare the relative attractiv
e-
nes
s of nearby stores. Furthermore, this technique
only

estimates the extents of trade area without the intermed
i-
ate inbound/outbound trajectories.

This study does not suggest

that the
Convenie
n-
ceProbe

system is superior to traditional consumer survey
method.
A t
raditional consumer survey can probe co
n-
sumers


attitudes and motivations
,
while
Convenie
n-
ceProbe

can only record true behaviors and movements.
However, in the case of collecting behavior patter
ns for
trade area analysis,
as in

the current study, using co
n-
sumers


own mobile phones can overcome the problems
of
imprecise

data
collection
due to human memory and
perception.

5.2
Mobile and UbiComp Technologies in Marketing
and Consumer Research

P
revio
us studies
focused on

how to augment shopping
experience
s

or promote the use of UbiComp technolog
ies

at stores.

Girgensohn

et al.

[
12
] use
d

fixed cameras to
monitor
the
in
-
store activities of retail shoppers and a
g-
gregate traffic flow of different store
sections. By analy
z-
ing
the
surveillance

videos captured from these cameras,
their

system detects and aggregates specific shopper
s’

activities

using heat maps
. This approach creates an

effe
c-
tive
method of
managing retail spaces.
Shopping Tracker
[
13
]
is

a p
hone
-
based shopping tracking system
that

mo
n-
itor
s

customers’ in
-
store shopping time. This system uses
a phone’s movement sensor to recognize unique sho
p-
ping movement trajectories resulting from store aisle la
y-
outs. Moiseeva
et al.

[10
] detect
ed

the
activit
ies of retail
shoppers, including their transportation modes and
in/out building
activities
, from their multi
-
day GPS tra
c-
es. They investigated the potential benefit of incorpora
t-
ing GPS, cell phone
,

and RFID technologies to reduce
the
self
-
reporting effor
t
s of

respondents

regarding

their
shopping trips. The collected pedestrian movement data
and consumer patterns provide insights
into
the retail
store location and customer time
-
space preferences.

To
enhance

the shopping experience,
Meschtscherj
a-
kov

et al.

[15
]
placed

a
dynamic

store map in a retail store.
This

store map combine
d

customer activity
visualization
(
e.g.
,

sales ranks) with traditional map elements (e.g.,
product locations).
Reitberger

et al.

[16
] set an interactive
mannequin in front of a shop window to
persuade

b
y-
passing customers to extend their time of stay. The pe
r-
suasive interactive mannequin
was

displayed on a large
LED screen and react
ed

to the presence of a user by alte
r-
ing body positions

and looking in their direction.
Kanda

et al.

[17
]
used a

social robot
to

distin
guish

potential cu
s-
tomers from passer
-
bys. Customer

s shopping
trajectories

were

recorded in a public space and used
to
identify
cu
s-
tomer profiles
and predict

potential custome
rs.

Decker
et
al.

[
23
]

implemented RFID
-
based
Smart Shelf technology

that tracks basic actions performed on items by customers,
such as
take
,
return
,

and
remove
. The data collected pr
o-
vides higher detail consumer in
-
store shopping behavior
tracking.

5.3
Ph
one
-
based Behavior Sensing Systems

Many researchers have used

phones to sense human a
c-
tivities. Madan
et al.

[1
1
]

leverag
ed a
phone’s Wi
-
Fi se
n-
sors, Bluetooth proximity sensors, and call logs to
record
social interaction and detect social patterns
.
They d
e-
signed a

phone
-
based sensing
platform

for
monitoring
epidemiological
huma
n behavioral patterns,
and found

individuals
exhibited

distinctive changes in behavior (e.g.
,

changes in frequency of communication) when
they

were
sick. With the ability to detect be
havior changes, the a
u-
thors stated the possibility of determining individual
health status and modeling epidemiological contagion
between people without medical heath reports.

Health related phone sensing projects aim to enhance
the health of users throu
g
h persuasion. Playful Bottle [
18
]
is
a
mobile persuasive system
that

encourages office
workers to drink healthy quantities of water. A mobile
phone attached to a mug detects the amount and regular
i-
ty of water consumed by the us
er. The UbiFit Garden sy
s-
tem
[
19
] encourages physical activities using on
-
body
sensing and personal displays on phones. UbiFit
displays

user exercise levels on a virtual flower garden shown on
the phone screen.

Several recent studies u
tiliz
e

mobile phones to sense

human

transportatio
n beha
vior. The UbiGreen project [
20
]

makes use of
the sensed transportation choice and di
s-
plays its environmental
effect

on a phone
. Focusing on
different aspects of transportation behavior, Zhen
g

et al.

[
14
] sens
ed

the transportation mode of users by classif
y-
ing and modeling
phone
-
collected
l
ocation data.

CenseMe [
21
] uses the microphone and accelerometer
of
a
mobile phone to infer the user’s activity an
d social
settings. SoundSense [
22
]
uses

the microphone to reco
rd
sounds from users’ daily lives. Machine Learning tec
h-
niques classify general sounds (e.g. music, voices) and
indentify novel sound events of special meaning to

the

users. Other projects tracked user location
-
visit history to
recommend be
tter routes or s
pots. GeoLife [
7
] is a friend
& location recommender system. Based on
the
human
trajector
y

traces recorded by GPS
-
enabled devices,
this

system models the location history of individuals and
identifies similarities among users. A user’s interests can
then b
e inferred and the system
can make

touring
spot/route and frien
d recommendations. Biketastic [
8
]
utilized the GPS sensors of smartphones to record cycling
trajectories and speeds.
Using

the accelerometer and m
i-
crophone sensors on the phone, this system
can

infer road
roughness and the general noise level along routes. Fina
l-
ly, each participant could provide his/her
biking
exper
i-
ences of the route
s

along with tags and descriptions while
riding.
Unlike

previous
studies
,
this

system expands the
breadth of
thes
e studies
by bringing mobile sensing into
customer behavior research.

AUTHOR ET AL.: TITL
E

11


C
ONCLUSION

This
study

presents a

novel application of phone
-
based
data collection to consumer behavior and marketing r
e-
search.
M
obile phones
were used
to outsource consumer
data collecti
on to everyday consumers, and to
automat
i-
cally
detect

the

behavior patterns
from their phones.
This
phone
-
based data collection system

was deployed

to co
l-
lect real customer flow data
from

neighborhood conve
n-
ience stores.
A

comparison

user study show
s

that it is
possible to use mobile phones to collect quality consumer
flow data and obtain accurate spatial patterns
. The results
of this approach are comparable

to

those from
the

trad
i-
tional face
-
to
-
face interview method. Given the ubiquity
of mobile phon
es, this
study

opens a door
to the

practical
use of phones from everyday consumers to sense and
report consumer behavior in marketing research. In other
words, this
study

goes beyond traditional
consumer self
-
reporting

to UbiComp’s
phone automated
-
reportin
g
, i.e., t
o-
ward an invisible data collection
method that does not
affect
natural consumer behavior.

The current ConvenienceProbe system has several li
m-
itations
,

which are also directions for future
research
.
First,
the

current data collection software only

runs on
several selected HTC Android mobile phones, which li
m-
its participation to those who have these HTC phones.
F
uture work
should

extend this data collection software
to other major phone platforms. Second,
the

current data
collection software stores
the collected sensor data on
phones and asks/requires user authorization prior to u
p-
loading each store visit data to the server.
F
uture work
should

con
sider
advanced

privacy mechanisms
.

Third,
the current GPS sensor errors create problems in custo
m-
er path/
trajectory reconstruction.
F
uture work
should

incorporate advanced post
-
processing techniques, such as
pedestrian walkways constraints, to improve
the
accuracy
of customer path/trajectory reconstruction.

R
EFERENCES

[1]

R.
.
J.

Kuo, S.C.

Chi, and S.S.

Kao
,


A
Decision Support System for
Selecting Convenience Store Location Through Integration of Fuzzy
AHP and Artificial Neural Network,
"

Computers in Industry,

vol.

47,
no.

2,
pp.
199
-
214
,
2002

[2]

M.

Levy, and B.A.

Weitz,
Retailing Management
,
7th Edition
.

McGraw
-
Hi
ll Higher Education, 2009.

[3]


Retailers Improve Service with Smart Security,
"

http://www.asmag.com/showpost/9586.asp
.

[4]

D.H.

Kim, Y.

Kim, D.

Estrin, and M.B.

Srivastava,

SensLoc: Sensing
Everyday Pla
ces and Paths using Less Energy
,”

Proc. ACM Conf
.
E
m-
bedded
Networked Sensor Systems

(Sensys ’
10
)
,
pp.
43
-
56
, 2010
.

[5]

D.L
.

Huff
,


A Probabilistic Analysis of Shopping Center Trade Ar
e-
as
,”

Land Economics
, vol.

39,

no.

1,
pp.

81
-
90
, 1963
.

[6]

B.
Gross
,


Dynamic Trade Areas. Business Geographics
,”

pp.

30
-
33
,
September
(1997)
.

[7]

Y.

Zheng, L.

Zhang, and
X.
Xie
,


Recommending Friends and Loc
a-
tions based on Individual Location History
,”

ACM Tran. Web,
vol.
5,

no.

1,
pp.

1
-
44
, 2011
.

[8]

S.
Reddy,
K.
Shilton,
G.
Denisov,
C.
Cenizal,
D.
Estrin, and
M.
Sr
i-
vastava
,


Biketastic:
Sensing and Mapping for Better Biking
,”

Proc.

ACM Conf
.

Human Factors in Computing Systems

(
CHI


10)
,
pp.

1817
-
1820
,

2010
.

[9]

A.R.
Patel,
T.J.
Fik, and
G.I.
Thrall
,


Direction Sensitive Wedge
-
casting for Trade Area Delineation
,”

J
.
Real Estate
Portfolio Management
,

vol.
14,
no.
2,
pp.
125
-
140
, 2008
.

[10]

A.
Moiseeva and
H.J.P.
Timmermans
,


Imputing Rele
-
vant Info
r-
mation from Multi
-
day GPS Tracers for Retail Planning and Ma
n-
agement using Data Fusion and Context
-
sensitive Learning
,”

J
.
Retai
l-
ing and
Consumer Services
,

vol.

17,
no.
3,
pp.

189
-
199
, 2010
.

[11]

A.
Madan,
M.
Cebrian,
D.
Lazer, and
A.

Pentland,

Social Sensing for
Epidemiological Behavior Change
,”

Proc. ACM Conf
.
Ubiquitous Co
m-
puting (Ubicomp


10)
, pp.

291
-
300
, 2010
.

[12]

A.
Girgensohn,
F.
Shipman, and
L.
Wilcox
,


Determining Activity
Patterns in Retail Spaces through Video Analysis
,”

Proc.

16
th

Conf.
Multimedia

(
MM

08
)
, pp.

889
-
892
, 2008
.

[13]

C.
-
W.
You,
C.
-
C.
Wei,
Y.
-
L.
Chen,
H.
-
H.
Chu, and
M.
-
S.
Chen
,


Shopping Time Monitoring in Physical
Stores Using Mobile
Phones
,”

IEEE Pervasive Computing
,
vol.
10,
no.
2,
pp. 37
-
43,
2011.

[14]

Y.
Zheng,
Y.
Chen,
Q.
Li,
X.
Xie, and
W.
-
Y.
Ma
,


Understanding
Transportation Modes Based on GPS Data for Web Application
,”

ACM Tran. Web
,
vol.
4,
no.
1,

pp.

1
-
36.

[15]

A.
Meschtscherjakov,
W.
Reitberger,
M.
Lankes, and
M.
Tscheligi
,

Enhanced Shopping: A Dynamic Map in a Retail Store
,”

Proc. ACM
Conf
.
Ubiquitous Computing (Ubicomp


08
)
,

pp.

336
-
339
, 2008
.

[16]

W.
Reitberger,
A.
Meschtscherjakov,
T.
Mirlacher,
T.
Scherndl,
H.
Huber, and
M.
Tscheligi
,


A Persuasive Interactive Mannequin for
Shop Windows
,”

Proc. fourth Conf. Persuasive Technology (Persu
a-
sive ’09)
, Article 4, 8 pages.

[17]

T.
Kanda,
D.
Glas,
M.
Shiomi,
H.
Ishiguro, and
N.
Hagi
-
ta
,


Who Will
Be the Customer? A Social
Robot that Anticipates People’s Behavior
from Their Trajectories
,”

Proc. ACM Conf
.
Ubiquitous Computing
(Ubicomp


08
)
,
pp.
380
-
389
, 2008.


[18]

M.
-
C.
Chiu,
S.
-
P.
Chang,
Y.
-
C.
Chang,
H.
-
H.
Chu,
C.
Chen,
F.
-
H.
Hsiao, and
J.
-
C.
Ko
,


Playful Bottle: a Mobile Social

Persuasion Sy
s-
tem to Motivate Healthy Water Intake
,”

Proc.
ACM Conf
.
Ubiquitous
Computing (Ubicomp
‘09
),

pp.
185
-
194.
, 2009.


[19]

S.
Consolvo,
D.W.
McDonald,
T.
Toscos,
M.Y.
Chen,
J.
Froehlich,
B.
Harrison,
P.
Klasnja,
A.
LaMarca,
L.
LeGrand,
R.
Libby,
I.
Smith, and
J.A.
Landay
,


Activity Sensing in the Wild: a Field Trial of Ubifit Ga
r-
den
,”

Proc.

ACM Conf
.

Human Factors in Computing Systems

(
CHI


08)
,
pp.

1797
-
1806
, 2008.

[20]

J.E.
Froehlich,
T.
Dillahunt,
P.
Klasnja,
J.
Mankoff,
S.
Consolvo,
B.
Harrison
,

and
J.A.
Landay
,


UbiGreen: Investigating a Mobile Tool
for Tracking and Support
-
ing Green Transportation Habits
,”

Proc.

ACM Conf
.

Human Factors in Computing Systems

(
CHI


09
)
,

pp.

1043
-
1052
, 2009.

[21]

E.
Miluzzo,
N.D.
Lane,
K.
Fodor,
R.
Peterson,
H.
Lu,
M.
Musole
si,
S.B.
Eisenman,
X.
Zheng, and
A.T.
Campbell
,


Sensing Meets Mobile S
o-
cial Networks: the Design, Implementation and Evaluation of the
CenceMe Application
,”

Proc. ACM Conf
.
Embedded Networked Sensor
Systems

(Sensys

08)
,
pp.
337
-
350
, 2008
.

[22]

H.
Lu,
W.
Pan,
N.D.
Lane,
T.
Choudhury
,

and
A.T.
Campbel
l,


SoundSense: Scalable Sound Sensing for People
-
Centric Sensing
Applications on Mobile Phones
,”
Proc.
7
th

Conf
.
Mobile Systems, Appl
i-
cations, and Services

(
MobiSys

0
9)
,
pp.
165
-
178
, 2009
.

[23]

C.
Decker,
U.
Kubach
,

and
M.
Beigl
,


Revealing the Retail Black Box
by Interaction Sensing
,”

Proc
.

23
rd

Conf
.
Distributed Computing Systems

(
ICDCSW

03
)
,
pp.
328
-
333
, 2003
.

[24]

T.J.
Stanley, and
M.A.
Sewall
,


Image Inputs to a Prob
abilistic Model:
Predicting Retail Potential
,”

J
.
Marketing
,
vol.
40,
no.
3,
pp.
48
-
53
, 1976
.