=Paper=
{{Paper
|id=Vol-522/paper-8
|storemode=property
|title=Short Paper: Situation-Awareness Model for Higher Order Network Knowledge Management Platform
|pdfUrl=https://ceur-ws.org/Vol-522/p4.pdf
|volume=Vol-522
|dblpUrl=https://dblp.org/rec/conf/semweb/MoonPK09
}}
==Short Paper: Situation-Awareness Model for Higher Order Network Knowledge Management Platform==
https://ceur-ws.org/Vol-522/p4.pdf
Short Paper: Situation-Awareness Model for Higher
Order Network Knowledge Management Platform
Aekyung Moon, Yoo-mi Park, Sang-gi Kim,
IP Convergence Technology Research Laboratory
Electronics and telecommunications Research Institute, Daejeon, Korea,
{akmoon, parkym, kimsang}@etri.re.kr
Abstract. In this paper, we propose the situation-awareness model for
HKMP(Higher order Knowledge Management Platform) that has a capability to
offer context-aware personalized services to user. HKMP is a platform that
provides the higher order knowledge from the contextual information of the
network and user ambient sensors through the knowledge processing techniques
including reasoning and learning. This paper presents the system architecture of
HKMP and classifies contextual information as lower order and higher order
knowledge. The proposed situation-awareness model for providing context-
aware personalized services recognizes the situation of the users
and recommends personalized services based on the information. The main idea
on this paper is how to evolve the awareness model without using personal
information causing privacy issues and how to draw an inference effectively
current situation of users. We continuously evolve our model to achieve this
requirement by the learning mechanism using the interaction between users and
mobile devices. As a result, we can make the user behavior pattern which can
be learned in situation and the situation is captured by union of sensors under
the current environments. In order to apply our model to new environments, we
simply need to define the sensor profiles without any change of model itself. So,
the proposed model consists of the pairs of context-action and deduce current
situation of users inference through the ontology model. At the end, we evaluate
the precision of the proposed approach through the use of Weka3 data mining
software with data sets of UCI machine learning depository. In the result of
evaluation, we expect HKMP to be an essential component to provide the
personalized services in the next generation networks.
Keywords: Network Knowledge, Ontology, Learning, Reasoning,
Recommender
1 Introduction
Since Mark Weiser proposed the concept of ubiquitous computing, a significant
amount of work has been devoted to context-aware [1-3]. Recently, various attempts
for providing the context-aware personalized services considering a situation as well
Proc. Semantic Sensor Networks 2009, page 110
as preference of a user are leading a new service paradigm of the next generation
networks. We therefore have developed the higher order knowledge management
platform (HKMP). The HKMP handles the network knowledge of the user
surrounding to provide context-aware personalized services that can be dynamically
adapted to a user’s current situation. To do this, HKMP gathers contextual
information from network as well as various sensors and classifies collected
contextual information as two categories: lower order and higher order knowledge for
handling efficiently it. Furthermore, HKMP provides higher order knowledge derived
from contextual information, which is helpful to create context-aware personalized
services according to user’s context or situation.
In this paper, we propose learner and ontology model for situation-awareness in
HKMP. To begin with, the proposed ontology model for situation-awareness defines
user centric lower order knowledge and their relationships including profiles, context,
and preferences of the user. Then, it derives user’s current situation using ontology
reasoning in order to providing users with context-aware personalized services. For
applying it to different domains, it consists of two types of ontologies; core part for
generic purpose and domain-specific part. We expect that it makes possible to
recommend personalized services at any given domain for the end users by deducing
simply user’s current situation using the proposed model.
Another key point is learner for situation-awareness. Our learner makes user
behavior patterns for recommending a weighted list of services to be preferred
according to user’s current situation. A several approaches have been proposed for
context-aware personalized services [4-6]. However, most of previous approach
assume static environment or fixed sets of context to determine user’s situation. To
overcome this problem, proposed model consists of context-action pairs and sensor
profile that can applicable to a variety of environment flexibly. For applying the
proposed model to new environment, we simply need to define the sensor profiles
without any changes of model itself. Moreover, it is important that learner is to evolve
model without the previous knowledge on a user. Therefore, we use reinforcement
learning that does not require external supervision. As a result, the HKMP learns user
behavior through both interaction with the user and user’s situation captured by
context. For the efficient normalization of context having continuous attributes, we
adopt minimal set of interval approach [7, 8]. We evaluate the precision of proposed
approach with data sets of UCI machine learning repository [9] to compare other
algorithms using Weka3 data mining software [10].
The remainder of this paper is organized as follows. Section 2 presents related
works. Section 3 discusses knowledge classification and system architecture of
HKMP. Section 4 proposes ontology and learning approach for situation-awareness
model of HKMP. Finally, section 5 summarizes this study.
2. Related Work
Personalization has become an important research area since the appearance of the
first papers on collaborative filtering in the mid-1990s [11]. There has been much
work done both in the industry and academia on developing new approaches to
Proc. Semantic Sensor Networks 2009, page 111
personalized recommender systems over the last decade. In the previous researches, it
has concentrated on contents recommendation associated with information retrieval
and filtering from the web. However, the current generation of personalization
requires further improvements to make advanced methods more effective and
applicable to an even broader range of real-life applications. Therefore, recent
personalization system deals with contextual information, user’s situation and
behavior patterns to improve user experiences [12]. However, existing models aren’t
sufficient to support situation-awareness. For example, CC/PP [13] as well as UAProf
[14] does not consider such information about the user’s behavior in case a specific
context or situation exists.
Although numerous definitions of situation-awareness have been proposed,
Endsley’s definition [15] is firmly established and widely accepted. Situation-
awareness involves being aware of what is happening around users. In this respect,
situation-awareness model contains information about the user’s behavior pattern and
user’s situations. This is to provide users with useful services now and in the near
future. There are several approaches for situation-awareness model, which assume
static environment or fixed sets of context to determine user’s situation [5,6]
The SPE (Secure Persona Exchange) [16] framework provides personalized
services to users in ubiquitous computing environments based on user preferences
stored on mobile devices; it does not fully include dynamic context information.
Daidalos [17] proposed four case of personalization in call redirection. Unlikely SPE,
Daidalos consider context, however, it use specific context defined in advance only.
MobiLife uses the learning mechanism to extract behavior patterns in the same
situation of the user or similar users [12]. In the neural network house, it is able to
predict occupancy of rooms and hot water usage using feedforward neural networks
[18]. Additionally, there are several approaches using learning [5,19]. Krause focused
on how machine learning techniques can identify typical user’s situation and modify
his mobile device’s settings based on experience [5]. Feki suggested the service
recommendation of robotics domain using Q-learning on user behavior [19].
3. Higher Order Knowledge Management Platform
3.1 Network Knowledge Classification
Lower Order Network Knowledge. The lower order knowledge includes all
information directly acquired from the underlying network. As shown in the Fig. 1,
the lower order knowledge consisted of information regarding the user and resources
(network, device, and service) that he/she could access, use, and are offered. We
defined profile as a collection of structured data that describes the static properties of
an object. Preferences are user’s conditional value of an object depending on context
and ambient information. Specially, we focused on service preference, which is a set
of information related to user’s preferred services, and service usage preference
acquired by learning mechanism. Based on the definition of context by Dey et al. [1],
context is any information that can be used to characterize the situation of an entity.
Proc. Semantic Sensor Networks 2009, page 112
We consider that sensed information from PCS (personal communication sphere),
current network bandwidth and user’s current location, etc.
Sensing PCS(Personal Communication Sphere)
Context info. preference
Schedule
Physical Terminal
presence
location status
current
Network
current BW network
access type
traffic
User’s network User’s device User’s service Preference
preference preference preference
User profile (SSid, Network Profile Service
Device Profile
age, gender, job) (Access Net., BW) Profile Profile
Service
Operator’s Repository
User Parlay X GW (TL, TS, Presence) Service Provider
information
IMS HSS Device
MLP Server IMS Presence
(physical
(physical
location, status,
server (activity, profile
location) mood, place-is) Device Vender
PUI, PRI, ..
Fig. 1 Network Knowledge Classification
Higher Order Network Knowledge. Higher order network knowledge could be
generated from lower order network knowledge using knowledge processing
technologies such as context reasoner, learner, and predictor in the Fig. 2. Those were
situation, intention, and pattern. Context reasoner has role to infer his/her situation
based on contextual information. The type of an actual situation, for example
“Waiting for bus”, can be recognized based on given contextual information.
Predictor is capable of predicting the services which the user is likely to want to use
in the near future. Suppose that he and his family arrive to international airport. And
they are on holiday, then predictor infers that their intention will “travel” and needs
“travel-related information” when arriving at the destination. Leaner has a capability
to extract rules and patterns out of massive usage history. Behavior pattern of a user
can be used in preferentially recommendation of service when the user is in the
specific situation. Recommender is to automatically identify the service categories to
be preferred in a given situation, user intention and user behavior pattern, etc.
3.2 System Architecture of HKMP
The system architecture of HKMP is shown in Fig. 3. Among these capabilities, the
core functions of HKMP for situation-awareness are context reasoner and learner.
Proposed context reasoner for situation-awareness has a capability to deduce user’s
current situation from available contextual information using ontology reasoning with
a predefined TBox schema. Leaner sets up and maintains service usage behavior
model using learning mechanism. In particular, the contextual information influences
user behavior model because it contains a pair of user behavior (service usage) and a
user situation consisted of contexts. The user service usage model is updated by user
pattern learner that analyzes user behavior history. The proposed model which
consists of {context, service} pair can be acquired by the context and the service
usage of a user; it then can be used to recommend personalized services according to
Proc. Semantic Sensor Networks 2009, page 113
user’s situation. We call these functions to SAM (Situation-Awareness Model) which
is organized by context reasoner and leaner for situation-awareness. The detailed
these algorithms are shown in section 4.
Recommender
Higher order Situation User Intention
User Behavior
Pattern
Knowledge
Context
Predictor Learner
Reasoner
Usage Rule/
Behavior Policy
User
Profile Context Network
Preference
Lower order Network Profile Preference
Profile Device
Knowledge Preference
Device
Profile Service
Terminal User PCS Environment Preference
Service Context Context Context Context
Profile
Fig. 2 Conceptual Architecture for HKMP
Context-aware
Context-aware Context-aware
Personalized
Personalized Service Personalized Service
Service
High order Knowledge Exposure
Recommender
Service Services Contents Contents
Info. Recommender Recommender Info.
Situation-Awareness Model
Context Ontology
Lower Reasoner Model Higher
order order
Know- Usage Pattern Know-
User Service
ledge Learner ledge
Usage Model
Knowledge Management Platform
Underlying Network (e.g. IMS)
Fig. 3 System Architecture of HKMP
4. Situation-Awareness Model
4.1 Situation-Awareness Ontology Model for Context Reasoner
Ontology Structuring. First of all, we drew competency question (CQ) lists which
could be asked for situation-awareness. After generalization of CQ, they were refined
Proc. Semantic Sensor Networks 2009, page 114
in the form of 4W (who, when, where, what) and 1H (how). Our basic CQ is “Which
service/device/network (what) is best when the user (who) is at (where) location or
situation”.
Based on the generalized CQ, we extracted the keywords such as user, service,
device, network, location, time, activity, and people. We found that every keyword
was closely related to the user. The user’s situation includes location of the user,
current time, schedule, location’s place type, etc. Therefore, we built user-centric core
ontology by representing a user and objects to be related the user as shown in Fig. 4.
Since situation and preference is domain-specific and especially situation is
characterized by derived contexts, we divide user-centric core and domain specific
parts for applying flexibly to different domains according to their purpose.
User Centric Service
Core Ontology
Network Device
Telematics domain
Home domain
Domain Specific
Ontology
User’s Preference User’s Situation Contexts
Meeting
Location
Service Shopping
Preference Place
Network … Type
Preference Situation
Preference
Schedule
WatchTV
Device
Preference …
Waiting
Time
ForBus
Telecommnucation domain
Fig. 4 Conceptual model for Situation-aware Ontology Structuring
Ontology Modeling. We depicted ontology model using OWL-DL in Figure 5.
Ontology model consisted of a user-centric core and domain-specific parts. The core
part is comprised of the Device, Service, User, Network, and Location classes. They
represent generic contexts to meet basic requirements of modeling for situation-
awareness. The domain specific part consisted of Preference, Group, Activity,
Schedule, and Presence. It was designed to provide personalized services in any given
domains. Preference is consisted of service preference and domain preference
modeled by N-ary relations of W3C.
In our ontology model, user’s situation is deduced from available information
using TBox predefined rules. We defined four cases of situation for
telecommunication domain as high-order knowledge; PersonInMeeting,
PersonInShopping, PersonInWatingForBus, and PersonInWatchingTV as shown in
Fig. 5. For example, ‘PersonInMeeting’ can be recognized based on the given context
such as location, role, schedule, and device status. Service ontology represents classes
for four service categories; Commerce, Information, Entertainment, Communication.
TV class is modeled as subclass of Entertainment and includes TV program genre
referenced by TV-anytime forum [19].
Proc. Semantic Sensor Networks 2009, page 115
Class hierarchy Object properties Implied relation Class Instance
User Centric Core Ontology Service Ontology
Thing Service
Category
Communica
uses Commerce tion
owns
Entertainment Information
Network Device User Service
owl:inverseOf TV
owl:inverseOf owl:inverseOf Music
Drama Action
hasAvailableNetwork isAvailableAt
hasLocation
locatedIn
isMemberOf Preference
hasPreference
hasPreference Domain
PersonIn Preference
Group
Shopping PersonIn Location Service
Watching Preference
Domain
TV Service
Waiting Preference1
PersonIn Situation Preference1
ForBus preferenceProbability
Meeting Hong
High
PlaceType Next Segment
Activity hasPreferenceValue
Schedule Service
Domain Specific Ontology Preference2
N-ary relations
PersonInMeeting ≡ User Π ∃hasActivity.Meeting Π ∃locatedIn.Office Π ∀own. DeviceInUse ⊆ User
DeviceInUse ≡ Device Π ∃isStatus.ON ⊆ Device
PersonInShopping ≡ User Π ∃hasActivity.Shopping Π(∃locatedIn.Arena∪∃locatedIn.Shopping-area∪ ∃locatedIn.Store)⊆ User
PersonInWaitingForBus ≡ User Π ∃locatedIn.BusStation ⊆ User
PersonWatchingTV = User Π ∃ has Activity.TV Π ∃locatedIn.Room ⊆ User
Fig. 5 Situation-aware Ontology Model for Context Reasoner of HKMP
4.2 Learner for Situation-Awareness
4.2.1 Basic Idea
Learner has the capability of keeping the situation model as patterns extracted from
the history of usage behavior. It can help to evolve user’s preference based on his
usage behavior patterns. Our learner analyzes user behavior history using feature of
reinforcement learning. The reinforcement learning method can proceed only through
interactions between each user and mobile devices without previously known
information. The representative studies using reinforcement learning are [20] and [21].
Our learner makes the user model by learning mechanism similar to reinforcement
learning, but has more advanced features. In order that our user model applies to new
environment, only sensor profiles are required to be defined without any modification
to model. Table 1 shows data structures of sensor profile. It defines that a set of states
s consists of contextual information as mentioned above. The user can perform any of
a set of possible action classes ac. The agent then receives a real-valued reward R.
Table 1 Data Structure of Sensor Profile
States = {c1, ..., cn}, 1≤ n, ck : kth context in the State
Attributes(ci) = {a i,1, ... a i,k}, 1≤ k and 1 ≤ i ≤ n
Action Classes = {ac1, ... acm}, 1≤ m
Reward R= { Selection-rs, Positive Feedback -rp, Negative Feedback-rn }
Proc. Semantic Sensor Networks 2009, page 116
Our learner has to two tables: context-aware user model (C-TBL) and a prediction
table (P-TBL). Suppose that the user’s current state s is comprised of three contexts
[user activity, location, time]. For each context, three C-TBL are needed. For example,
if a user gets up in the morning and listens to news (ac1) in his or her bedroom, this
becomes the [c1: wakeup, c2: bedroom, c3: morning] situation. C-TBL[c1][ac1], C-
TBL[c2][ac1] and C-TBL[c3][ac1] are updated by learner. P-TBL is designed to predict
and recommend. To recommend user action in current situation, P-TBL is computed
by recommender using the C-TBL. The action classes are the services provided by the
service provider. That is, the value stored in the P-TBL[cs][ack] shows the preference
of action ack in the corresponding current situation(cs). The user directly chooses an
action and the next state is determined according to the selected action. Therefore, in
the proposed scheme, we modify the rule (1), because the reward value for a
recommendation is affected by the behavior information between the system and user.
The learning equation is given as:
Q( s, a) ← Q( s, a) + γR, R ∈ {rs , rp , rn }
(1)
The detailed algorithm of the learning phase is as follows. In the Step 1, If C-TBL
for ck(ck ∈ States) doesn’t exist, create C-TBL for context ck using context profiles.
The context profile is required to register attribute values which each context has.
Refer sensor profile information in the Table 1. If C-TBL for ck (ck ∈ States) doesn’t
exist, create C-TBL for context ck using sensor profiles.
Step 2 is initialization phase for new context ck. Initialize new C-TBL for ck, set 0
to C-TBL[ak,i][acj], for each ak,i ∈ Attributes(ck), acj∈Action Classes. Initialize value
of R for R ∈ {rs, rp, rn}.
Step 3 repeats the following steps.
Step 3-1: Input current situation s(t), s(t) is consisted of sum of ak,i(t), where ak,i(t)
∈ Attributes(ck) and k ∈ {1, ..., n}. The relationship among state, current state and
current situation is as follows:
∀ck, ck ∈ Current States ⇒ ck∈States
Situation(s(t), x) ⇔ ∀x, x ∈ Attribute (ck), and ck∈ Current States
Step 3-2: if ai,k(t) is continuous value, it needs to be normalization. We adopt simple
normalization approach and minimal set of interval approach for discretization of
context with continuous attributes.
Case 1. min-max normalization performs a linear transformation on the original
data and makes fixed intervals. Suppose that mina and maxa are the minimum and the
maximum values of ai,k(t).
ai,k− mina
ai(,tk) = ai',k = (new_ maxa −new_ mina ) + new_ mina
maxa − mina
Case 2. we quantize continuous value into k intervals, where k is the minimal
constant depending on the range of particular feature. In order to discretize feature A
into k intervals, we use basic heuristic [7] to find a minimal set of intervals. Context
data sets with continuous values will fall into an appropriate space
Step3-3: Input an current action ac(t) by user selection. Determine R(t) according to
user behavior information. Update the C-TBL as following rules:
for each ck in C-TBL[ak,i(t)][ac(t)] do
C-TBL[ak,i][ac(t)]← C-TBL[ak,i(t)][ac(t)]+αR(t),
Proc. Semantic Sensor Networks 2009, page 117
where α is the discount factor and ck∈ States.
To provide the individualized and active services by using the learned C-TBL
tables, P-TBL can determine whether it is a best to recommend any kind of action and
recommends the action to a user in the particular state.
∑
P − TBL[ci ][ac ] = M (cs) w × C − TBL[a ][ac ] (2),
k
ai ∈cs
i i k
where M(cs) is a normalization term, wi represents the weight of ci, and cs is user’s
current situation. wi represents the weight of ci and can be set identically. However,
the significance of each context can vary with a user. For example, there is a sensitive
user on the illumination sensor compared to other peoples. In order to differentiate the
significance of each contexts interpreted by sensors, the entropy of a context is
calculated through the method that is proposed in [8]. Then, preference for an action
ack, Pref(ack) can be estimated by the following rule using P-TBL:
Pr ef (ack ) = ∑ P − TBL[ai ][ack ]
ai ∈cs
4.3. Evaluation
Learner must be evaluated by real users in the ubiquitous environments. However, it
is difficult to acquire behavior pattern of the real user; because of some critical
problems related to the privacy issue of individuals. Keskustalo [22] noted that
experiments on the effectiveness of relevance feedback with real users are time-
consuming and expensive; therefore, leaner has been tested on several sets of data in
UCI depository. To evaluate operation of leaner, we chose the following the data set;
Iris, Wine, Create Approval, Balance and balloon in UCI repository. For example, Iris
is perhaps the best known database to be found in the pattern recognition literature. In
the case of the create approval in UCI machine learning repository datasets [9], this
consists of 15 contexts which has 9 categorical attributes and 6 continuous attributes,
and two action classes as show in Table 2.
Table 2. Example of UCI Datasets for Simulation
Data Instance Attr.(Categorical) ActionClass
Create Approval 665 15(9) 2
Balloons 20 4(4) 2
Balance 625 4(4) 3
Iris 150 4(0) 3
Wine 178 13(0) 3
And also, we chose machine learning algorithms from the Weka3 data mining
software [10]: J48, ZeroR, NaiveBayes [23] and SMOSupport Vector Machine. J48
builds a C4.5 decision tree Naïve Bayses selects the most likely classification based
on a set of attribute values using prior probabilities and conditional densities of
individual features. ZeroR simply predicts the majority class in categorical or average
class if the class is numeric [23]. The k-fold cross validation is used in order to raise
the confidence of experiments. The performance evaluation metric in this experiment
Proc. Semantic Sensor Networks 2009, page 118
is the accuracy (precision). When R is being the number recommended as a user and
the RP (Recommended Preference) is being the number which a user actually prefers,
precision is calculated as the RP / R and showed by the %. In order to raise the
confidence of experiments, k-fold cross validation is used. K-fold cross validation is
one way to improve over the holdout method. The data set is divided into k subsets,
and the holdout method is repeated k times. However, when computing by equation
(2), the user preference by the ontology was not considered in order to experiment in
same condition with the comparison algorithm. wi is calculated by the entropy of
context through the information gain [8]. We implemented two algorithms: Learner-S
and Learner-Q. Learner-S uses the min-max normalization and quantizes continuous
value into 10 fixed intervals. On the other hand, Learner –Q quantize continuous
value into minimal k intervals. And we only consider rS because that it is difficult for
deciding the reward value without explicit user feedback.
The left side of Fig. 6 shows the precision of each algorithm according to
categorical context only. Learner is better than other algorithms in the aspect of
Create Approval. The precision of our learner is 86.8% at create approval. And in all
data sets, our learner is better than ZeroR with 1.5 times. The right side of Fig. 6
shows the precision of each algorithm according to continuous context only. Learner-
Q is better than other algorithms in the aspect of Iris and Wine data sets. In the case of
Wine data sets, Learner-Q improves 2.4times compared to ZeroR even though
improves slightly compared to J48. Fig. 5 shows that Learner-Q is better than
Learner-S. In the case of Wine data sets, Learner-Q improves with 1.2 times
compared to Learner-S.
create Approval Balance balloon iris wine
120 120
100 100
80 80
60
60
40
40
20
20
0
0 Learner-S Learner-Q J48 ZeroR Naïve SMO
Leaner J48 ZeroR Na? e Bayses SMO Bayses
Fig. 6 Results of Evaluation
5. Conclusions
We propose the situation-awareness model for HKMP(Higher order Knowledge
Management Platform) that has a capability to offer context-aware personalized
services to user. This paper presents the system architecture of HKMP and classifies
contextual information as lower order and higher order knowledge. The Proposed
Proc. Semantic Sensor Networks 2009, page 119
situation-awareness model is aware of user’s situation and recommends personalized
services based on this information. The main idea on this paper is how to evolve the
awareness model without using personal information causing privacy issues and how
to draw an inference effectively current situation of users. To achieve this requirement,
we evolve our model through interactions between users and mobile devices using
learning mechanism. And it derives user’s current situation using ontology reasoning.
Moreover, we adopt minimal set of interval approach for improving performance
about discretization of context having continuous attributes.
We evaluated the precision of proposed approach using Weka3 software with data
sets of UCI machine learning depository. The precision of our learner is 86.8% at
create approval data set. And in all data sets, our learner is better than ZeroR
algorithm with 1.5 times. Learner-Q is better than other algorithms in the aspect of
Iris and Wine data sets. In the case of Wine data sets, Learner-Q improves 2.4times
compared to ZeroR even though improves slightly compared to J48. In the case of
Wine data sets, Learner-Q improves with 1.2 times compared to Learner-S. For
further study, we have a plan to provide wholly implementation of HKMP approaches
includes higher order knowledge exposure layer.
Acknowledgments. This research is supported by the IT R&D program of
MKE/IITA of South Korea. [2009-F-048-01, Development of Customer Oriented
Convergent Service Common Platform Technology based on Network].
References
1. A. Dey, G. Abowd and D. Salber, “A Conceptual Framework and a Toolkit for
Supporting the Rapid Prototyping of Context-aware Applications,” Human Computer
Interaction, 16, (2001)
2. H. Kranenburg, M. Bargh, S. Lacob and A. Peddemors, “A Context Management
Framework for Supporting Context-Aware Distributed Applications,” IEEE
Communications Magazine, pp. 67-74, (2006)
3. A. Moon, H. Kim, H. Kim and S. Lee, “Context-Aware Active Services in Ubiquitous
Computing Environments,” ETRI Journal vol. 29, no. 2, pp. 169-178, (2007)
4. M. Sutter, O. Droegehorn and K. David, “User Profile Management on Service Platforms
for Ubiquitous Computing Environment,” IEEE Vehicular Technology Conf., pp. 287-
291, (2007)
5. A. Krause and A. Smailagic, “Context-Aware Mobile Computing: Learning Context-
Dependent Personal Preferences from a Wearable Sensor Array,” IEEE Trans. On Mobile
Computing, 5(2), pp. 113-127, (2006)
6. J. Fogarty et al., “Predicting Human Interruptibility with Sensors,” ACM Trans. On
Compute-Human Interaction, 12(1), 119-146, (2005)
7. S. Nguyen and A. Skowron, “Quantization of Real Value Attributes,” Joint Annual Conf.
Information on Sciences, pp. 34-37, (1995)
8. K.Jearanaitanakij and O. Pinngern, “An Information Gain Technique for Acceleration of
Convergence of Artificial Neural Networks,” ICICS, pp. 349-352, (2005)
9. UCI Machine Learning Repository, “http://archive.ics.uci.edu/ml/”
10. Weka3 Data mining Software, http://www.cs.waikato.ac.nz/ml/weka/
Proc. Semantic Sensor Networks 2009, page 120
11. G. Adomavicius, “Towards the Next Generation of Recommender Systems: A Survey of
the State-of-the-Art and Possible Extensions,” IEEE trans. On Knowledge and Data
Engineering, 17(6), pp. 734-749, (2005)
12. IST-2004-511607, MobiLife D27b (D4.1b) v1.0, (2004)
13. W3C, Composite Capabilities/Preference Profiles (CC/PP), (2004)
14. OMA, Use Agent Profile (UAProf) v2.0, (2003)
15. M. Endsley, “Toward a theory of situation awareness in dynamic systems,. Human
Factors 37(1), pp. 32-64, (1995)
16. A. Brar and J. Kay, “Privacy and Security in Ubiquitous Personalized Applications,” User
Modelimg Workshop on Privacy-Enhanced Personalization, (2005)
17. S. McBurney, M. Williams, N. Taylor and E. Papadopoulou, “Managing User Preference
for Personalization in a Pervasive Service Environment,” IEEE Conf. on
Telecommunications (AICT'07), (2007)
18. M. C. Mozer, “The Neural Network House: An Environment that Adapts to Its
Inhabitants,” AAAI, pp. 110–114, (1998)
19. M. Feki, S. Lee, Z. Bien and M. Mokhtai, “Context Aware Life Pattern Prediction Using
Fuzzy-State Q-Learning,” LNCS 4541, pp.185-195, (2007)
20. The TV-Anytime Forum, Specification Series: S-3 On: Metadata, (2001)
21. F. Herndex, E. Gaudioso and J. Boticario, “A Reinforcement Learning Approach to
Achieve Unobtrusive and Interactive Recommendation Systems for Web-Based
Communities,” AH 2004, LNCS3137, pp. 409-412, (2004)
22. H.Keskustalo, Kalervo and A. Prikola, “The Effects of Relevance Feedback Quallity and
Quantity in Interactive Relevance Feedback: A Simulation Based on User Modeling,”
ECIR 2006, LNCS 3936, 2006, pp. 191-204, (2006)
23. S. Louis, A.Shankar, “Context Learning Can Improve User Interaction, Information
Reuse and Integration,” IEEE conf. IRI, pp. 115-120, (2004)
Proc. Semantic Sensor Networks 2009, page 121