A Multi-domain Hybrid Recommender Systems Based on a Dynamic

Contextual Ontological User Proﬁle

Aleksandra Karpus and Krzysztof Goczyła

Faculty of Electronics, Telecommunications and Informatics, Gda´nsk University of Technology,

ul. Gabriela Narutowicza 11/12, 80-233 Gda´nsk-Wrzeszcz, Poland

1 STAGE OF THE RESEARCH

The aim of research presented here is creation of

a multi-domain hybrid recommender system based on

a dynamic contextual ontological user proﬁle. Re-

cently, we have built a contextual ontology for rep-

resenting user preferences. The next step that need to

be undertaken is validation of correctness and com-

pleteness of this idea of representing a user proﬁle.

We consider three context parameters: location,

time and user mood. The validity of these parame-

ters, and hence, their impact on user preferences, has

been conﬁrmed by the results of a survey among users

of recommendation systems. The research on knowl-

edge aquisition and new recommendation algorithms

is still in the early stages.

2 RESEARCH PROBLEM

The area of recommender systems is well established

but scientists and engineers still try to improve quality

and diversity of recommendations. Some of research

questions in this ﬁeld that still can be addressed are:

• How can system automatically aquire and contin-

uously improve user proﬁle?

• What impact on the recommendations quality has

a user context?

• How to append a contextual data to the user pro-

ﬁle?

During our research we will try to answer those and

some additional questions:

• Does the way of describing user preferences in the

form of contextual ontology signiﬁcantly improve

the quality of recommendations in comparison to

known methods?

• How do we measure similarity of two contextual

ontologies?

• Does removal of some concepts reduce the possi-

bilities of reasoning from contextual ontology?

3 OUTLINE OF OBJECTIVES

The objectivesof presented research is to developnew

methods and algorithms in the ﬁeld of recommender

systems. Our main goal is novel approach to model-

ing user preferences. We plan to derive an algorithm

for user preferences aquisition. This algorithm will

create a new proﬁle ontology based on the gained in-

formation and a ,,metaontology” from which needed

concepts and roles would be extracted. Due to the new

representation of a user proﬁle, new recommendation

algorithms need to be developed. The method of com-

paring two contextual ontologies is also needed.

4 STATE OF THE ART

The role of recommender systems is to give possibly

the most adequate recommendations to users in differ-

ent situations. Scientists try to include a user context

in recommendation algorithms to give more adequate

results to users (Adomavicius and Tuzhilin, 2011).

There are several defﬁnitions of what the context is.

In the rest of this paper as a context we understand

”any information that can be used to characterize the

situation of a person, place or object that is considered

relevant to the interaction between a user and applica-

tion” (Blefari-Melazzi et al., 2007).

Another trend in the ﬁeld of recommender sys-

tems is to apply ontologies to capture user prefer-

ences. It has been proved that ontological user proﬁle

improves recommendation accuracy and diversity (Su

et al., 2012).

Recently, some context-aware recommender sys-

tems using ontologies for making recomendations

have been developed and reported. AMAYA rec-

ommender system that allows management of con-

textual preferences and contextual recommendations

was proposed (R¨ack et al., 2006). AMAYA also uses

an ontology-based content categorization scheme for

mapping user preferences to entities to recommend.

Karpus A. and Goczyla K..

A Multi-domain Hybrid Recommender Systems Based on a Dynamic Contextual Ontological User Proﬁle.

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

News@hand, a hybrid recommender system for

ontology-based personalized and context-aware rec-

ommendations of news items was presented in (Can-

tador et al., 2008). The news items are automatically

and periodically retrieved via RSS feeds and anno-

tated with semantic concepts from system domain on-

tologies. During an interaction with the user a set of

weighted concepts from the domain ontologies is col-

lected. A user context is represented by this set. The

importance of concepts fades away with the time by a

decay factor. This helps to keep the user context up to

date. Existing relations between concepts in the on-

tologies are used to ﬁnd semantic paths linking pref-

erences to context.

A context-aware system which recommends Web

services to users was described in (Rodriguez et al.,

2013). The main idea is using a multi-dimensional

ontology model to describe Web services, a user

context and an application domain. The multi-

dimensional ontology model consists of a three inde-

pendent ontologies: a user context ontology, a Web

service ontology and an application domain ontology,

which are combined into one ontology by some re-

lations between concepts from different ontologies.

Data from a WSDL ﬁle for a Web service are auto-

matically added into the Web service ontology during

the registration process. The user must specify the

name, birth date, sex and occupation to build his pro-

ﬁle. The user context ontology consists of those pa-

rameters and a list of interests. Every item of the list

has a level of interest property, which is used to as-

sign a weight to the item during the recommendation

process.

5 METHODOLOGY

In this section we describe an idea of a contextual ap-

proach to an ontological user proﬁle (Subsection 5.2).

First, we need to deﬁne the contextualontology which

is done in Subsecion 5.1. The general architecture of

the proposed system is presented in Subsection 5.3.

A description of planned experiments and future re-

search close this section.

5.1 Structured-interpretation Model

Contextual ontology introduced in (Goczyła et al.,

2007) enables us to model different situations in

which a user could ﬁnd himself as a set of ontological

modules. As an ontology we mean here a Description

Logics (DL) ontology which consists of a terminol-

ogy (TBox) and a world description (ABox). As a

context we understand a part of TBox deﬁned by val-

A A

T T

5 6 7 8

specializes

context

contextinstance

instance-of

is-aggregated-by

Figure 1: Structured-Interpretation Model (from (Goczyła

et al., 2007)).

ues of a set of contextual parameters. The contexts are

arranged in an inheritance hierarchy. More special-

ized terminologies may ,,see” more general ones, but

more general terminologies are unaware of the exis-

tence of more specialized ones. To deal with possible

different assertional parts of the knowledge base we

can create many ABoxes for one terminology. These

ABoxes are called context instances. To allow for

a ﬂow of conclusions between context instances, the

context instances are connected by relation of aggre-

gation. This approach to modularization, described in

details in (Goczyła et al., 2012), is called Structured-

Interpretation Model (SIM) and is illustrated in Fig. 1.

To explain how the contextual ontology works we

use a simple example taken from (Waloszek, 2010).

Let assume that an ontology consists of two contexts

(TBoxes) and three context instances (ABoxes). Con-

text T

provides concept Can

resuscitate from which

concept Doctor inherits. Doctor is a concept pro-

vided in the terminology of context T

. Context in-

stances A

and A

describe a situation of an individ-

ual called john

doe from different points of view: he

is a doctor in Poland but legally he is not a doctor

in United Kingdom. Assertions Doctor( john

doe)

and ¬Doctor( john doe) are contradictory, neverthe-

less the ontology is consistent. It is because the con-

cept Doctor is deﬁned below the context instance A

which aggregates context instances A

and A

. The

concept Can

resuscitate is immanent for the context

because it is deﬁned on the level of the context in-

stance A

. Therefore the conclusions reﬂecting the

fact that John Doe can resuscitate can ﬂow between

the two contexts (in Poland and in the United King-

dom). This example is shown in Fig. 2.

Another example is shown in Fig. 3. Here we have

three contexts: T

that describes general notions of

Woman and Man, T

that specializes T

towards de-

scription of voices in a choir, and T

that also spe-

cializes T

but towards description of social relations.

IC3K2014-DoctoralConsortium

:Can_resuscitatem¨

:Doctor Can_resuscitatem

:Doctor(john_doe)

:¬Doctor(john_doe)

:-

Poland

UnitedKingdom

general

Can_resuscitate(john_doe)

Figure 2: An example of SIM ontology (based on

(Waloszek, 2010)).

:Man Woman* m¦

:Soprano Womanm

:Soprano(mary)

:Married(mary)

:-

Contextvillechoir

Contextvilleregisteroffice

Contextvilleataglance

Woman(mary)

Wife(mary)

:Wife Woman Marriedm *

Figure 3: An example of SIM ontology (from (Goczyła

et al., 2012)).

Context instance A

aggregates context instances A

and A

. Although ABox of A

is empty, interpretation

ℑ

in order to be a model of the knowledgebase has to

assign Mary to the concept Woman. As a consequence

of this, the same rule enforces that in the interpreta-

tion ℑ

Mary is assigned to the concept Wife, as the

information about Mary being a woman “ﬂows” down

the aggregation relationships.” (Goczyła et al., 2012).

As we could see from the above examples, an on-

tology built according to SIM enables us to model

complicated structure of user preferences depending

on the context in which the user currently is.

5.2 Contextual Approach to Ontological

User Proﬁle

A user proﬁle keeps all the information about user

preferences. For assuring context-awareness in our

recommender system, it is crucial to embed contex-

tual parameters into a user proﬁle. To this aim, we

propose a new representation of a user proﬁle as a

contextual ontology.

We consider three contextual parameters: loca-

tion, time and mood - which divides a proﬁle ontol-

ogy into context modules. By location we mean the

place and other circumstances that affect the user ac-

tivity, in particular: Is he at work? Or at home? Or

on holidays? It is important to know if the user is at

home or at work to make appropriate recommenda-

A A

5 6

specializes

context

contextinstance

instance-of

is-aggregated-by

8 9

general

work

home

time

mood

time=morning

time=weekday

morning

time=

night

time=weekend

mood=

sad

mood=

happy

Figure 4: The general idea of a contextual user proﬁle on-

tology (based on (Goczyła et al., 2007)).

tions. For example, assume that we want to recom-

mend some books to a Java developer who is know to

have a daughter. We do not want to recommend him

a book with fairy tales when he is at work, but we

want to do that when he is at home. Instead, we want

to recommend him a book about a new framework in

Java when he is at work.

The location parameter is a special parameter in

our approach. Different values for this parameter re-

quire different terminological parts, thus we limited

possible values to the following two: at home and at

work. It is shown in Fig. 4.

The time parameter is considered in two dimen-

sions. The ﬁrst one speciﬁes if the current day is

a weekday or a weekend day. The second one speci-

ﬁes the part of day: morning, afternoon, evening and

night. Such a division is justiﬁed by the fact that nor-

mally we are searching for different things on the In-

ternet on weekday morning (for example, some news

update) than on weekend evening (for example, what

movies are shown in nearby cinemas).

Another contextual parameter that is worth adding

into the user proﬁle in user mood. The fact that one is

happy or sad could really inﬂuence his behavior and

current preferences. Let us consider a simple exam-

ple. Mary had fallen in love and actually she is very

happy. She is a great fan of documentary movies.

If we did not take into account her current mood,

we might recommend her an interesting movie about

the Second World War. But Mary might not want to

watch this movie at this time because it is too depress-

ing for her.

AMulti-domainHybridRecommenderSystemsBasedonaDynamicContextualOntologicalUserProfile

Wi-Fi Facebook Callendar

Contextualuserprofile

Recommendationalgorithm

user

Figure 5: The general architecture of the context-aware rec-

ommender system based on an ontological user proﬁle.

Initially the proﬁle ontology will not contain any

context instances. When we ﬁrst get information

about a user (i.e. when he registers to the system),

new ontology will be created. The terminological part

of it will contain only those concepts and roles from

the model taxonomy (,,metaontology”) that would be

relevant to the data retrieved and stored as the asser-

tional part of the proﬁle ontology.

5.3 The Architecture of the System

One of the main problems in the ﬁeld of recommender

systems is the cold-start problem. Our system will

be partly based on collaborative ﬁltering techniques,

thus it should be exposed to the existence of this prob-

lem. On the other hand, we model user preferences by

using a contextual ontology, thus we use advantages

of the semantic approach to modelling. To avoid the

cold-start problem, we will provide to our system a

possibility to log in with a speciﬁc Facebook account.

Then, after a user logs into the system for the ﬁrst

time, we will automatically build a proﬁle ontology

based on the user data obtained from the social media

account. To make this process transparent to the user,

we would do it in background and for making recom-

mendations ,,in real time” after his ﬁrst login we use

only the user’s latest ,,likes”. It would give us possi-

bility to get access to a lot of user data that he shares

throughout the social media. In case the user does not

intend to use the Facebook account, he could create a

new one and give us only some selected information

about himself, including e.g. his current mood.

The current position of the user we could get from

GPS or a Wi-Fi adapter. The current time we will ob-

tain from the system clock and the calendar. To get the

information about a user location (at work or at home)

we need to know the user’s schedule. The simplest

way to do that is to connect our recommender with

the Google Callendar, assuming that most of users al-

ready use it.

If the user shares with the system information

about his current mood, the system could return the

most appropriate recommendation - the recommenda-

tions will be made basing on the most speciﬁc context

(bottom nodes of the graph in Fig. 4). If not, the rec-

ommendation will be performed in more general con-

text in the user proﬁle ontology. For this purpose we

need to compare a speciﬁed context of the user X with

the same contexts in different users proﬁles to ﬁnd

a k-Nearest Neighbors. After that, we could recom-

mend items to the user X basing on the preferences

of top k-Nearest Neighbors in that speciﬁc context.

The general architecture of the context-aware recom-

mender system based on the ontological user proﬁle

is shown in Fig. 5.

5.4 Planned Experiments and Research

The next step that needs to be undertaken is validation

of correctness and completeness of the proposed user

proﬁle. To this aim we plan to conduct qualitative re-

search in a group of some ﬁfty people. Firstly, we will

create some kind of subsystem which will allow only

to register, log in, and present to the user data from the

proﬁle broken down into contexts. The people will be

divided into two groups. The ﬁrst one will be asked to

should create new account from scratch. After that, all

people will complete a survey about their proﬁle in-

formation and corectness of dividing their preferences

into contexts.

Next, we will conduct at least two experiments.

The ﬁrst one will involve testing the inﬂuence of the

size of the ontology (and richness of hierarchy) on the

time efﬁciency and the quality of recommendations.

The second one will cover comparison of the results

returned by the proposed system with the results re-

turned by existing recommendation systems.

6 EXPECTED OUTCOME

The aim of the decribed research project is to provide

a new multi-domain context-aware recommender sys-

tem that will:

• minimalize the cold-start problem;

• improve the quality and diversity of recommenda-

tions;

• facilitate search of relevant information from In-

ternet that would be of interest to a user;

IC3K2014-DoctoralConsortium

• automatically create an adequate user proﬁle;

• allow for dynamic update of a user proﬁle.

To achieve the results, some new methods and al-

gorithms in the area of recommender systems will be

developed.

REFERENCES

Adomavicius, G. and Tuzhilin, A. (2011). Handbook on

Recommender Systems, chapter Context-Aware Rec-

ommender Systems, pages 217–256. Springer.

Blefari-Melazzi, N., Casalicchio, E., and Salsano, S.

(2007). Context-aware service discovery in mobile

heterogenous environments. In Proc. IEEE Mobile

and Wireless Communications Summit, pages 1–5.

Cantador, I., Bellogin, A., and Castells, P. (2008).

Ontology-based personalised and context-aware rec-

ommendations of news items. In Web Intelligence and

Intelligent Agent Technology, Proc. WI-IAT ’08, pages

562–565.

Goczyła, K., Waloszek, A., and Waloszek, W. (2007). Con-

textualization of a dl knowledge base. Proc. of the

20th International Workshop on Description Logics

DL07, pages 291–298.

Goczyła, K., Waloszek, A., Waloszek, W., and Zawadzka,

T. (2012). Modularized knowledge bases using con-

texts, conglomerates and a query language. Intelligent

Tools for Building a Scientiﬁc Information Platform,

390:179–201.

R¨ack, C., Arbanowski, S., and Steglich, S. (2006). Context-

aware, ontology-based recommendations. In Inter-

national Symposium on Applications and the Internet

Workshops (SAINTW 2006), pages 98–104.

Rodriguez, J., Bravo, M., and Guzman, R. (2013). Multi-

dimensional ontology model to support context-aware

systems. In AAAI 2013 Workshop.

Su, Z., Yan, J., Ling, H., and Chen, H. (2012). Research

on personalized recommendation algorithm based on

ontological user interest model. Journal of Computa-

tional Information Systems, 8(1):169–181.

Waloszek, A. (2010). Hierarchiczna kontekstualizacja baz

wiedzy. PhD thesis, Gda´nsk University of Technology.

AMulti-domainHybridRecommenderSystemsBasedonaDynamicContextualOntologicalUserProfile