This article was downloaded by: [Aston University]On: 28 August 2014, At: 12:02Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH,UK

Cybernetics and Systems: AnInternational JournalPublication details, including instructions forauthors and subscription information:http://www.tandfonline.com/loi/ucbs20

CONTEXTUALIZEDRECOMMENDATION BASED ONREALITY MINING FROM MOBILESUBSCRIBERSJason J. Jung a , Hojin Lee b & Kwang Sun Choi ca Department of Computer Engineering , YeungnamUniversity , Gyeongsan, South Koreab KTF , Seoul, South Koreac Saltlux , Seoul, South KoreaPublished online: 06 Feb 2009.

To cite this article: Jason J. Jung , Hojin Lee & Kwang Sun Choi (2009)CONTEXTUALIZED RECOMMENDATION BASED ON REALITY MINING FROM MOBILESUBSCRIBERS, Cybernetics and Systems: An International Journal, 40:2, 160-175, DOI:10.1080/01969720802634089

To link to this article: http://dx.doi.org/10.1080/01969720802634089

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all theinformation (the “Content”) contained in the publications on our platform.However, Taylor & Francis, our agents, and our licensors make norepresentations or warranties whatsoever as to the accuracy, completeness,or suitability for any purpose of the Content. Any opinions and viewsexpressed in this publication are the opinions and views of the authors, andare not the views of or endorsed by Taylor & Francis. The accuracy of theContent should not be relied upon and should be independently verified withprimary sources of information. Taylor and Francis shall not be liable for any

losses, actions, claims, proceedings, demands, costs, expenses, damages,and other liabilities whatsoever or howsoever caused arising directly orindirectly in connection with, in relation to or arising out of the use of theContent.

This article may be used for research, teaching, and private study purposes.Any substantial or systematic reproduction, redistribution, reselling, loan,sub-licensing, systematic supply, or distribution in any form to anyone isexpressly forbidden. Terms & Conditions of access and use can be found athttp://www.tandfonline.com/page/terms-and-conditions

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

CONTEXTUALIZED RECOMMENDATION

BASED ON REALITY MINING FROM MOBILE

SUBSCRIBERS

JASON J. JUNG1, HOJIN LEE2, andKWANG SUN CHOI3

1Department of Computer Engineering, YeungnamUniversity, Gyeongsan, South Korea2KTF, Seoul, South Korea3Saltlux, Seoul, South Korea

It is difficult to be aware of the personal context for providing a

mobile recommendation, because each person’s activities and

preferences are ambiguous and depend upon numerous unknown

factors. In order to solve this problem, we have focused on a reality

mining to discover social relationships (e.g., family, friends, etc.)

between people in the real world. We have assumed that the personal

context for any given person is interrelated with those of other

people, and we have investigated how to take into account a person’s

neighbor’s contexts, which possibly have an important influence on

his or her personal context. This requires that given a dataset, we

have to discover the hidden social networks which express the con-

textual dependencies between people. In this paper, we propose a

semiautomatic approach to build meaningful social networks by

repeating interactions with human experts. In this research project,

we have applied the proposed system to discover the social networks

among mobile subscribers. We have collected and analyzed a dataset

of approximately two million people.

This work was supported by the IT R&D program of MIC=IITA. (2007-S-048-02,

Development of Elementary Technologies for Fixed Mobile IP Multimedia Convergence

Services on Enhanced 3G Network).

Address correspondence to Jason J. Jung, Department of Computer Engineering,

Yeungnam University, Dae-Dong, Gyeongsan Republic of Korea 712-749, South Korea.

E-mail: j2jung@gmail.com, horilla@ktf.com, or kschoi@saltlux.com

Cybernetics and Systems: An International Journal, 40: 160–175

Copyright Q 2009 Taylor & Francis Group, LLC

ISSN: 0196-9722 print=1087-6553 online

DOI: 10.1080/01969720802634089

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

INTRODUCTION

Context-awareness services for the mobile computing community have

been heavily studied to provide useful content and information to mobile

subscribers anytime and anywhere. There has been lots of work to

extract, represent, and reason a variety of pieces of contextual infor-

mation detected from the mobile computing environment. Researchers

have basically been trying to recognize meaningful relationships between

user actions and a certain combination of contexts, which is regarded as

a set of condition-action rules. Eventually, they believe that if a user is

under a certain condition and the condition is realized to be matched

to the predefined contexts, the consequent rule has to be conducted

for him or her.

There are many contexts that must be taken into account, including

not only physical contexts (e.g., spatial and temporal contexts) of envi-

ronments, but also cognitive contexts (e.g., preferences, mental states,

and social affinities) of human users. Such context models are the

5W1H (Abowd and Mynatt 2000), the stochastic segment model (Strang

and Linnhoff-popien 2004), and the ontology-based context model

(Gu et al. 2004; Jung 2008a).

It is still difficult, however, for service providers to be aware of

personal context at a certain moment and place, due to a number of

uncertainties. These unpredictable factors can be dealt with by several

contextual fusion approaches (Sekkas et al. 2007), which are intended

to integrate as many contexts as possible. Such an approach has been

applied in various domains such as location-based systems (Cabri et al.

2003; Korpipaa et al. 2003), as well as multiple expert systems (Herrera

and Martinez 2001), image processing by multitemporal contextual

information (Melgani et al. 2001), and information retrieval systems

(Pant and Srinivasan 2006; Jung 2007).

In this work, we expect to improve the performance of contextual ser-

vices provision (i.e., minimizing the error of predicting personal contexts)

by discovering social networks. The important assumption behind these

approaches is that contexts on user-behaviors are interrelated with others

they are socially related to. Also, the stronger the social relations between

people, the more dependent and influenced is their context by other

contexts. Recently, referring to the work of Mollenhorst et al. (2008) in

social network communities, we have realized that the context of a certain

user is strongly dependent on those of his acquaintances (e.g., families,

REALITY MINING 161

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

colleagues, and friends). It means that a user can make a different decision

and take a different action under the same environmental and cognitive

contexts, depending on whom he is currently staying with.

For example, in Table 1, even though Paul is in a restaurant (i.e.,

Case1 and Case2), his context might differ (i.e., Ctx1 and Ctx2) according

to his guests (e.g., his father or his girlfriend). Sometimes the personal

context of a user is significantly influenced by socially-related people,

and vice versa (Irving and Conrath 1988; Cross et al. 2001). Hence, it

is referred to as a social affinity context in this paper.

The next issue to be aware of is the social affinity context, which

reveals a meaningful social network from mobile users. We have

employed the reality mining method to reveal the real social network

from online medium communication. In particular, Eagle and Pentland

(2006) have reported their experiment by using various sensors. Unless

we ask users to describe their own social networks, it is difficult to make

sure whether two arbitrary users share a social affinity in common. In

this study, data mining tools have been employed to statistically analyze

log datasets and limited profile information. More importantly, social

network ontology is exploited to logically construct a given social

network to discover hidden information.

In this paper we want to introduce a novel approach to the

interactive social network discovery method. Basically, interactive prob-

lem solving methodologies have been regarded as an efficient way to find

better solutions for dealing with very complex problems. For instance, as

human users can interact with Web search engines (i.e., modify and

adjust their query terms over time), they can eventually search for

more relevant information from the Web. This process is quite similar

to the well-known generalization and specialization processes on version

spaces (Hong and Tseng 1997).

More particularly, in regard to organizing social networks among

mobile subscribers, it is a difficult and time-consuming task to find out

which conditions between two users can be derived from various social

Table 1. An example of context dependencies of Paul

Contexts Spatial context Social affinity context Integrated context

Case1 Restaurant Girlfriend Ctx1

Case2 Restaurant Father Ctx2

Case3 Hospital Father Ctx3

162 J. J. JUNG ET AL.

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

relations. Here, we can say that the query terms on Web search engines

are contextualized with the conditions represented as a propositional

sentence. While the interactions for this query transformation are simply

based on a human user’s intuitions, the proposed interactive systems in

this paper can automatically return statistical results and supplementary

information by applying data mining modules to collected usage patterns.

The deductive information for a certain social relation is referred to as

social propositions and is suggested to experts for help in decision-making

processes. Thus, as another solution to improve the performance of

context-based service, we have been motivated to build a social network

among mobile users in this research project. It is a research project called

Next-generation Intelligent Content-delivery Enabler (NICE) for deliver-

ing personalized information to mobile devices via the social networks.

Real customer information has been acquired from KT Freetel (KTF),

one of the major telecommunication companies in Korea.

The remainder of this paper is organized as follows: we will first

describe the problem on reality mining (i.e., building social networks) from

mobile communications. Then, more importantly, the contextual depen-

dency by social network ontology will be explained. Following this we will

explain the interactive discovery of social networks and show how to pro-

vide context-based services by using social networks among mobile users,

respectively. We will then mention some related work on personalization

and building social network between mobile users. Finally, we will draw a

conclusion and address the ongoing and future work of this project.

CONTEXTUAL DEPENDENCY FROM SOCIAL AFFINITY

To establish better personalized services based on personal contexts, we

have to take into account contextual dependency, i.e., some contexts can

be changed by being influenced by other contexts. Mainly, we are focus-

ing on social affinity contexts. The assumption is that the personal

context of a person is dependent on the contexts of others, if and only

if the person is socially connected to them. In other words, we can say

that the context CxtA of user UA is depended on (i.e., able to be changed)

CxtB of user UB, if user UA is socially connected with UB.

The main benefit of contextual dependency is that the user activities

are assumed to be restricted. In other words, we can more easily predict

which actions will be taken by the user under the contexts by extracting

contextual dependencies from social affinities.

REALITY MINING 163

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

For example, in Figure 1, if we suppose that Michael is the father of

Paul (i.e., <Michael, isFatherOf, Paul>), then we claim a following context:

Michael’s context CxtðMichael;LocationÞ ¼ ‘‘Department store’’ on Paul’s

context CxtðPaul; DateÞ ¼ ‘‘Birthday’’ throughsocial affinity ð‘‘isFatherOf ’’Þ:

When tomorrow is Paul’s birthday, and Michael is in a department store,

he might be trying to buy some presents for his son.

REALITY MINING BASED ON ONTOLOGY

Thereby, in terms of reality mining, we have to discover useful social net-

works from a collected dataset. For doing this, we have to analyze and

exploit two parts of information; (1) a user-activity dataset collected

from a number of sensors monitoring the corresponding people for a

while, and (2) social network ontology which is a background knowledge

indicating real social relations. By boosting both of them, we eventually

want to find out as many social propositions as possible.

In this section, we want to introduce an iterative approach to clarify

and establish the social networks. By the divide-and-conquer method, a

multiplex social network is separated and superposed with respect to the

contexts (Jung et al. 2007). In addition, the social network ontology is

playing a role of indirectly inferring social relationships between two

persons. For example, in Figure 1, the social relationship ‘‘isGrand-

Father’’ between Michael and Tomas is inferred by combining the social

relationships between Michael and Paul and between Tomas and Paul.

Basically, a social network is represented as a graph-structured network.

Definition 1 (Social network) A social network S is defined as

S ¼ hV ;Ri

where V and R are sets of participants fv1; . . . ; vjVjg, and relation-

ships between the participants represented as an adjacency matrix,

respectively.

Figure 1. An example of social networks.

164 J. J. JUNG ET AL.

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

In particular, R is not necessarily symmetric, because we want to

consider the directionality of the relations. Of cause, the links can be

weighed for representing the strength of social ties between two users.

In this work, the social network can contain multiple contexts together

at the same time, because the proposed system allows several different

ontologies to be imported.

Definition 2 (Multiplex social network) A multiplex social network Sþ is

defined as

Sþ ¼ hV ;R;Ci

where V and R are the same components as normal social networks S.

Additionally, component C is able to specify social affinity contexts of R.

A simple example of a multiplex social network is shown in Figure 1.

Links in this multiplex social network are attached with other different

social affinities (i.e., C¼ {isFatherOf, isSonOf, isWorkingWith}) all together.

Furthermore, in this paper we focus on a social network in which

two arbitrary persons are linked with more than one link. The link

between two persons is attached with some concepts to describe the

affinities between them. We assume that the concepts applied to label

the links by using semantics derive from several different social network

ontologies (described in next section), e.g., family ontology, company

ontology, and university ontology. Thus, this semantically multiplex

social network Sþ is represented as

Sþ ¼ hV ;R;L;CRi

where V ¼ fv1; v2; . . . ; vjV jg is a set of nodes (i.e., actors participating in

S), and R � V � V is a set of links (or edges) representing relations

between actors. L is a finite set of labels defined by actors participating

in the social network. This label can more specifically describe the con-

text of social relationships (i.e., R) between two actors. CR � R� L is a

set of associating multiple labels attached to each relation rij 2 R. As a

matter of fact, these labels are supposed to be derived with concepts

in social network ontologies. Thus, the multiple labels of a relation

rij 2 R between vi and vj are represented as a set of triples

fhvi; vj ; cxi j cx 2 CR; x 2 ½1;Xij �gwhere the relation is connected from vi to vj, and Xij is the number of con-

cepts labeled to the link rij. We assume that labels cx are contextualized

by retrieving semantics from the social network ontology of vi or vj

(or somehow, both of them). The direction of links determines which

REALITY MINING 165

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

social network ontology is applied to which label. It means that cx is a

semantic substructure of vi’s ontology, because rij is built by source actor

vi to target actor vj.

Social Network Ontology

We have investigated a social network ontology. It is to support the logi-

cal inference for finding out the relationships between social affinities

from a given multiplex social network. As shown in Figure 1, the indirect

social relationship ‘‘isGrandFather’’ between Michael and Tomas is

inferred by the social network ontology. The ontology is containing

relevant knowledge, e.g., isFatherOf¼ isSonOf 1.

Therefore, in this work, we want to employ the ontology which pro-

vides semantics to label social affinities on social networks. The ontology

alignment method can be applied to measure the semantic distance

(meaning the complementary to similarity) between the corresponding

contexts of social affinities. Furthermore, as clustering the equivalent

social affinities, consensual ontologies are discovered to support seman-

tic bridging among social network ontologies.

For aligning ontologies, we exploit the similarity measurement strategy

(Euzenat and Valtchev 2004; Jung and Euzenat 2007), which defines all

possible similarities (e.g., SimC, SimR, SimA) between classes, relationships,

attributes, and instances. For simplicity, we want to consider only SimC.

Given a pair of classes (i.e., c 2 O and c 2 O 0) from two different social

network ontologies O and O ’, the similarity measure SimC is assigned by

SimCðc;c0Þ¼X

F2N ðCÞpC

F MSimY FðcÞ;Fðc0Þð Þ¼pCL simL LðcÞ;Lðc0Þð Þ

þpCsupMSimC F supðcÞ;Fsupðc0Þð ÞþpC

subMSimC FsubðcÞ;Fsubðc0Þ� �

where L is the label of a class and N (C)¼fFsup;Fsubg is the set of all rela-

tionships (e.g., subclass and superclass) in which class C participates. The

weights pCF are normalized (i.e.,

PF2N ðCÞp

CF ¼1). Also, the set functions

MsimC compute the similarities of two entity collections. The best alignment

between two ontologies O and O 0 can be established by finding a maximal

matching maximizing the summed similarity between the features:

MSimC O;O0ð Þ¼max

Phc;c0i2Pairingðo;o0ÞSimCðc;c0Þ

� �

maxðjOj; jO0jÞ

166 J. J. JUNG ET AL.

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

in which function Pairing provides a match for the two set of classes.

Methods like the Hungarian method allow us to directly find the pairing

which can maximize the similarity. As a result, we can obtain a set of

semantic correspondences between two ontologies.

For example, in Figure 2, from two ontologies, class ‘‘Children’’ is a

superclass of ‘‘Son’’ and ‘‘Daughter’’, and the relation ‘‘isSonOf ’’ is

inversed with ‘‘isFatherOf.’’

NETWORK SEPARATION AND SUPERPOSITION

The contextual dependencies from multiplex social networks are found

out by the two phases, which are based on the divide-and-conquer

approach. This function is needed to help users improve their under-

standability about a large-scale social network. It means that users can

be aware of topological patterns of the network as well as semantic rela-

tionships between two arbitrary persons within a network. Eventually,

this information plays an important role in reality mining for contextual

dependencies.

During the divide step, we separate a given multiplex social network

Sþ. With respect to concepts c, which is labeling social links r2R, we can

divide S into a set of subnetworks. A subnetwork for a concept cm is

represented as

shmi ¼ hVm;Rm;L; cmi

Figure 2. Alignment between social network ontologies.

REALITY MINING 167

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

where Vm � V ;Rm � Vm � Vm. It seems rather simple and trivial, but, more

importantly, the social network experts can easily understand local internal

structural patterns P (shki) of a large-scale, multiplex social network. Such

patterns not only involve centrality, betweenness, and diameter of a social

network, but also semantic centrality (Jung 2008b).

For the conquer step, a set of subnetworks separated in the divide

step should be superposed reversely. Thus, we have to combine the local

social features (or patterns) P (shki) discovered from a subnetwork with

others. The overall social features are formulated by computing the sum-

mation of local social patterns of users, as sequentially mapping two sub-

networks in the set of subnetworks. Here, we have to focus on measuring

the similarities between ck and the other labels in order to discover the

semantically equivalent (or relatively close) concepts on multiplex social

network S. To find the semantic correspondences, we need to exploit the

semantic alignment function by measuring the similarity between con-

cepts, which is introduced in the previous section. After each mapping,

we have to conduct normalization. This phase can estimate the indirect

social relationships between any two arbitrary persons by composing the

concepts labelling the links. Of course, this is only to suggest possible

contextual dependencies to the experts for reality mining.

INTERACTIVE DISCOVERY SYSTEM FOR SOCIAL NETWORKS

In this paper, we have investigated interactive discovery process for

analyzing a large amount of datasets, including usage patterns collected

from mobile users. The main goal of this discovery process is to enrich

the social propositions from a certain multiplex social network. The

main steps for the interactive discovery process involve the following:

Step 1: a human expert inputs a propositional sentence without any

quantitative modifier to the system.

Step 2: the proposition is translated into mathematical algebra by

referring to the social network ontology. This step involves

symbols to be compared.

Step 3: a data mining module can scan the collected dataset to measure

the confidence level.

Step 4: the confidence level of the proposition is shown to the expert.

These steps are repeated until the best combination of conditions are

found, as shown in Figure 3.

168 J. J. JUNG ET AL.

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

Now, we want to explain about the mobile usage pattern datasets

sampled from KTF legacy databases where raw records are stored.

Mainly, it consists of three parts: (1) registration profiles, (2) device

(and service) specifications, and (3) the calling patterns of over 60,000

mobile subscribers. These datasets are applied to predict social relations

between mobile users by our heuristics. We have tried to formalize the

scenarios which are easily understandable. Thus, each scenario can be

matched to a set of social relations. Given a certain social relation, we

have investigated as many cases as possible. To do so, we have built a

decision tree by interviewing domain experts in KTF and by using

machine learning software packages (e.g., SPSS Clementine [http://

www.spss.com/clementine/] and Weka [http://www.cs.waikato.ac.nz/

ml/weak/]). This process can find out the best orders of fields to verify

the scenarios.

We believe that the social propositions can be discovered (i.e.,

adjusted and modified) as interacting with the proposed system. There

are two types of enriching social propositions. First, by adjusting a vari-

able, each social relationship should be described by as precise social

propositions as possible. For example, suppose that we have to find

out the social propositions of a social relation ‘‘isFatherOf.’’ During

Step 1 and 2, an expert can assert the following propositional conditions

between two users A and B:

. P1: both of last names are equivalent (�).

. P2: the difference between both ages is more than 20

In step 3, both social propositions defined by the expert are evalu-

ated by statistical software packages (i.e., data miner), and in step 4,

Figure 3. Interactive discovery process for social propositions.

REALITY MINING 169

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

an expert can find out that the confidence values of L (P1, �) and L (P2,

20) are 0.99 and 0.67, respectively. Here, this process can be iteratively

executed, according to the expert’s decision. If he or she wants to

improve the confidence values of the second condition P2, it might be

modified as ‘‘the difference between both ages is more than 25,’’ and

L (P2, 25) make the confidence level 0.78.

Second, social propositions between social relationships should be

described. We want to give a simple example of two social relations

‘‘isFatherOf ’’ and ‘‘isFamilyWith,’’ which are the most important social

relation in this project, from two different ontologies. In step 1 and 2

we can say that nA is a family with nB, when either P1 or P2 in Table 2

is satisfied.

By ontology alignment, we know that ‘‘isFatherOf ’’ is a subclass of

‘‘isFamilyWith’’ and we can specify the social propositions of isFamilyWith

to acquire those of isFatherOf. During interacting with the proposed data

mining module, the real dataset can be analyzed to annotate each prop-

osition with statistical confidence. Thus, the proposition from Table 2

can be branched to more specific propositions, as shown in Table 3.

In addition, these logical expressions can be dynamically updated

over time. More importantly, the social relations are semantically repre-

sented as concept hierarchy of social network ontologies. For example,

Table 2. Social Propositions by Common Sense About isFamilyWith;

Operator ^ Means Logical Conjunction

P1: (Payment(nB)¼ nA)

^ (Lastname (nA)¼Lastname(nB))

P2: (Location (nA, AtNight)¼Location (nB, AtNight))

^ (Lastname (nA)¼Lastname (nB))

Table 3. Adjusted Social Propositions by Interactions About isFatherOf

P1–1: (Payment (nB)¼ nA, 60%)

^ (Age (nA) – Age (nB)2 [30, 50], 0.85)

^ (Lastname (nA)¼Lastname (nB))

P1–2: (Payment (nB)¼ nA, 60%)

^ (Age (nA) – Age (nB)2 [40, 50], 0.60)

^ ( Lastname (nA)¼Lastname (nB))

P2: (Location (nA, AtNight)¼Location (nB, AtNight))

^ (Age (nA) – Age (nB)2 [30, 50])

^ (Lastname (nA)¼Lastname (nB))

170 J. J. JUNG ET AL.

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

‘‘isFatherOf ’’ and ‘‘isBrotherOf ’’ are subclasses of ‘‘isFamilyWith.’’ Thus,

when the information given by users is unclear or ambiguous, we can

replace it to one of its super-class relations.

CONTEXT-BASED SERVICE

To efficiently support personalized service, various types of information

can be applied for modelling a target user’s preference. One well-known

approach, the so-called collaborative filtering (Konstan et al. 1997;

Pazzani 1999), compares profiles of people. Such profiles are composed

of ages, genders, occupation, and so on. The main assumption of this

approach is that the closer profile should be the more like-minded people.

It means that two persons whose ages are the same are more probably inter-

ested in the same movie, compared to people who are older (or younger).

However, in the real world, current personalized services have not

shown efficient performance; people are not satisfied with these services

at all. We think that most of the personalization mechanisms are trying to

uncover hierarchical clustering structures (this is very similar to the

decision tree), which identify cohesive user-groups, of which members

in the same group might have more in common with each other (e.g.,

movies and music) than with other groups (Herlocker et al. 2000). This

statistical analysis to extract simple demographic features by comparing

user profiles (e.g., ages, genders, and so on) identify personal context—

what they are looking for in a certain situation (Bonhard et al. 2006). In

other words, the personal recommendation for each user is supposed to

be more specific.

In order to solve this problem, we mainly take into account two more

conditions. First, social affinity is regarded as a reasonable evidence to

predict the personal context. For example, social relations (particularly,

kin relations and friendships) can be assumed to identify each person’s

context more specifically. When a person is looking for his or her father’s

birthday present, it is much more probable that he or she is looking for

what his or her father wants, than what the person doing the search

wants. In aggregating the social networks, we can build a social network

for making various social relations extractable. This social network can

play a role of integrating multiple contexts which are interrelated

between each other (e.g., parents and children).

The second condition is location, that is, the geographical position

where one is. This personal context changes over time. In order to

REALITY MINING 171

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

support better personalization, the service should be timely activated.

For example, a user who is in a department store is expected to buy a

certain product, rather than to go to a restaurant.

More importantly, these two types of conditions can be merged for

better personalized recommendation. Given two persons who are linked

(highly related) on a social network and located in (or moving to) a

close place, we can recommend very reasonable information to them.

Especially, in this research project, we have been focusing on the mobile

users joining KTF services. The problem is that the social networks are

hidden. Therefore, we want to discover the hidden social network from

usage patterns of mobile devices.

RELATED WORK AND DISCUSSION

Kazienko (2007) proposed a way of measuring the closeness between two

persons to build a social network that focused on the calling patterns,

that is, when and how frequently people make a call. Eagle and Pentland

(2006), in their ‘‘reality mining’’ project, introduced experimental results

for evaluating several hypothesis. These hypotheses have been compared

to self-reports provided by users.

Personalization based on multiagent systems has been introduced in

MAPIS (Petit-Roze and Strugeon 2006). With regard to the business-

oriented work in Changchien et al. (2004), the personalization process

on e-commerce has been conducted by three modules; (1) marketing

strategies, (2) promotion patterns model, and (3) personalized pro-

motion products. Especially, location-based personalization services

ave been implemented, as in NAMA (Kwon et al. 2005).

CONCLUDING REMARKS AND FUTURE WORK

This work is an on-going research project for delivering personalized con-

tent to mobile users. In this paper we have introduced our interactive

approach to construct meaningful social networks between mobile users.

Each context included in a social network has been combined with a

spatial context to produce a better recommendation.

For future work, we are planning to put the calling patterns into

the social network which has been built by expert heuristics. It will

make the social network more robust and dynamically evolvable.

Furthermore, the evaluation method has been considered to verify

whether the personalized service is reasonable or not.

172 J. J. JUNG ET AL.

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

REFERENCES

Abowd, G. D. and Mynatt, E. D. 2000. Charting past, present, and future

research in ubiquitous computing. ACM Transactions on Computer-Human

Interaction 7(1): 29–58.

Bonhard, P., Harries, C., McCarthy, J., and Sasse, M. A. 2006. Accounting for

taste: Using profile similarity to improve recommender systems. Proceedings

of the SIGCHI Conference on Human Factors in Computing Systems

(CHI 2006), New York: ACM, 1057–1066.

Cabri, G., Leonardi, L., Mamei, M., and Zambonelli, F. 2003. Location-

dependent services for mobile users. IEEE Transactions on Systems, Man,

and Cybernetics–Part A: Systems and Humans 33(6): 667–681.

Changchien, S. W., Lee, C.-F., and Hsu, Y.-J. 2004. On-line personalized sales

promotion in electronic commerce. Expert Systems with Applications 27(1):

35–52.

Cross, R., Rice, R. E., and Parker, A. 2001. Information seeking in social

context: Structural influences and receipt of information benefits. IEEE

Transactions on Systems, Man, and Cybernetics–Part C: Applications and

Reviews 31(4): 438–448.

Eagle, N. and Pentland, A. 2006. Reality mining: Sensing complex social systems.

Personal and Ubiquitous Computing 10(4): 255–268.

Euzenat, J. and Valtchev, P. 2004. Similarity-based ontology alignment in OWL-

Lite. Proceedings of the Sixteenth European Conference on Artificial Intelligence

(ECAI 2004), Valencia, Spain: IOS Press, August 22–27, 333–337.

Gu, T., Wang, X. H., Pung, H. K., and Zhang, D. Q. 2004. An ontology-based

context model in intelligent environments. Proceedings of International

Communication Networks and Distributed Systems Modeling and Simulation

Conference (CNDS 2004), 101–106.

Herlocker, J. L., Konstan, J. A., and Riedl, J. 2000. Explaining collaborative

filtering recommendations. Proceedings of the 2000 ACM Conference on

Computer Supported Cooperative Work (CSCW 2000), New York: ACM,

241–250.

Herrera, F. and Martinez, L. 2001. A model based on linguistic 2-tuples for dealing

with multigranular hierarchical linguistic contexts in multi-expert decision-

making. IEEE Transaction on Systems, Man, and Cybernetics–Part B:

Cybernetics 31(2): 227–234.

Hong, T.-P. and Tseng, S.-S. 1997. A generalized version space learning

algorithm for noisy and uncertain data. IEEE Transactions on Knowledge

and Data Engineering 9(2): 336–340.

Irving, R. H. and Conrath, D. W. 1988. The social context of multiperson,

multiattribute decisionmaking. IEEE Transactions on Systems, Man and

Cybernetics 18(3): 348–357.

REALITY MINING 173

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

Jung, J. J. 2007. Ontological framework based on contextual mediation

for collaborative information retrieval. Information Retrieval 10(1):

85–109.

Jung, J. J. 2008a. Ontology-based context synchronization for ad-hoc social

collaborations. Knowledge-Based Systems 21(7): 573–580.

Jung, J. J. 2008b. Query transformation based on semantic centrality in semantic

social network. Journal of Universal Computer Science 14(7): 1031–1047.

Jung, J. J. and Euzenat, J. 2007. Towards semantic social networks. Proceedings of

the 4th European Semantic Web Conference (ESWC 2007), Lecture Notes in

Computer Science. Innsbruck: Springer, 4519: 267–280.

Jung, J. J., Juszczyszyn, K., and Nguyen, N. T. 2007. Centrality measurement

on semantically multiplex social networks: Divide-and-conquer approach.

International Journal of Intelligent Information and Database Systems 1(3=4):

277–292.

Kazienko, P. 2007. Expansion of telecommunication social networks. Proceedings

of the 4th International Conference on Cooperative Design, Visualization, and

Engineering (CDVE 2007), Lecture Notes in Computer Science. Shanghai,

China: Springer, 4674: 404–412.

Kleinberg, J. M. 2001. Small-world phenomena and the dynamics of information.

Advances in Neural Information Processing Systems, Vancouver, British

Columbia, Canada: MIT Press, 14: 431–438.

Konstan, J. A., Miller, B. N., Maltz, D., Herlocker, J. L., Gordon, L. R., and

Riedl, J. 1997. Grouplens: Applying collaborative filtering to usenet news.

Communications of the ACM 40(3): 77–87.

Korpipaa, P., Mantyjarvi, J., Kela, J., Keranen, H., and Malm, E.-J. 2003.

Managing context information in mobile devices. IEEE Pervasive Computing

2(3): 42–51.

Kwon, O., Choi, S., and Park, G. 2005. NAMA: A context-aware multi-agent

based web service approach to proactive need identification for personalized

reminder systems. Expert Systems with Applications 29(1): 17–32.

Melgani, F., Serpico, S. B., and Vernazza, G. 2001. Fusion of multitemporal con-

textual information by neural networks for multisensor image classification.

Proceedings of the 2001 IEEE International Geoscience and Remote Sensing

Symposium (IGARSS 2001), 7: 2952–2954.

Pant, G. and Srinivasan, P. 2006. Link contexts in classifier-guided topical

crawlers. IEEE Transactions on Knowledge and Data Engineering 18(1):

107–122.

Pazzani, M. J. 1999. A framework for collaborative, content-based and

demographic filtering. Artificial Intelligence Review 13(5–6): 393–408.

Petit-Roze, C. and Strugeon, E. G.-L. 2006. MAPIS: A multi-agent system

for information personalization. Information & Software Technology 48(2):

107–120.

174 J. J. JUNG ET AL.

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4

Sekkas, O., Anagnostopoulos, C. B., and Hadjiefthymiades, S. 2007. Context

fusion through imprecise reasoning. Proceedings of the 2007 IEEE

International Conference on Pervasive Services (ICPS), 88–91.

Strang, T. and Linnhoff-popien, C. 2004. A context modeling survey. Proceedings

of the Workshop on Advanced Context Modelling, Reasoning and Management

at The Sixth International Conference on Ubiquitous Computing

(UbiComp 2004), 234–241.

REALITY MINING 175

Dow

nloa

ded

by [

Ast

on U

nive

rsity

] at

12:

02 2

8 A

ugus

t 201

4