Using Hypergraph-based User Proﬁle in a Recommendation System

Hilal Tarakci and Nihan Kesim Cicekli

Department of Computer Engineering, Middle East Technical University, Ankara, Turkey

Keywords:

User Modeling, User Proﬁle, Hypergraph-based User Model, Knowledge Representation, Recommendation

System.

Abstract:

We propose a hypergraph-based user proﬁle which facilitates aggregating partial proﬁles of the individual and

obtain a complete, multi-domain user model. The aggregation involves a semantic enhancement procedure

which results in enriched user proﬁles. The proposed user model is capable of extracting general and domain-

based user proﬁles and answering several connected data queries such as recommendation, in reasonable time.

In this paper, we present a recommendation case study which uses the proposed user model and illustrate the

traversal algorithms for a variety of connected data problems.

1 INTRODUCTION

The popularity of social networking sites has dramat-

ically increased over the last decade. The user’s pro-

ﬁle can be extracted by examining the individual’s

behavior (Gauch et al., 2007). The user’s activities

on social websites reveal important information about

his/her proﬁle. Social networks differ in nature and

are used for various purposes. For instance Face-

book is used for social interaction and entertainment

whereas LinkedIn is used only for professional in-

terests. Therefore, mining separate social networks

independently results in partial proﬁles of the user

which merely represent user’s interests for one or few

domains. Seamless aggregation of partial user pro-

ﬁles obtained from different knowledge sources is still

an unsolved problem. In this paper, we present the

implementation of a hypergraph-based user model to

aggregate partial proﬁles of the individual to obtain

a complete, semantically enriched, multi-domain user

model and show that it can be used for different rec-

ommendation purposes.

The employed user proﬁle structure is mutually

associated with the aggregation methodology. The

aggregation process depends on the predeﬁned user

model data structure, and this structure is deﬁned

according to the main goals of the aggregation. If

the main purpose is producing an interoperable user

model, the proﬁle is generally deﬁned by a standard

(Orlandi et al., 2012) or user-deﬁned (Wischenbart

et al., 2012; Ghosh and Dekhil, 2008) ontology. In

this paper, one of our main goals is solving con-

nected data problems such as recommendation effort-

lessly. An effective solution strategy for connected

data problems is matching an entrance point to the

data structure and traversing the neighbours accord-

ing to the speciﬁed algorithm. Therefore, graphs nat-

urally support connected data problems (Robinson

et al., 2013). The vertices usually represent the items

and the users where an edge between a user and an

item indicate user’s interest on that item. The edges

could be associated with weights which represent the

strength of the relation between the vertices. Since

the graph is only capable of representing binary rela-

tions, other approaches have been proposed for han-

dling higher order relations in user modelling domain.

There are a few studies which deﬁne user model as

bipartite (Tiroshi et al., 2013) and tripartite graphs

(Chen et al., 2012). In general, if the number of vertex

types n is known in advance and the relations in the

user model are binary, an n-partite graph is capable of

representing the proﬁle. However, if there are higher-

order relations, a hypergraph is more appropriate to

represent the user model (Li and Li, 2013; Kramar

et al., 2013; Bu et al., 2010).

In a previous paper we presented the initial ideas

for using hypergraph in the modelling of user pro-

ﬁles (Tarakci and Cicekli, 2012a; Tarakci and Cicekli,

2012b). In this paper, the main contributions are (i)

a user proﬁle representation with hypergraphs, (ii) a

semantic aggregation methodology and (iii) a recom-

mendation case study to illustrate the solution for var-

ious connected data problems.

The paper is organized as follows. Section 2 sum-

marizes the related studies. Section 3 deﬁnes the pro-

posed hypergraph based user model and presents the

Tarakci H. and Kesim Cicekli N..

Using Hypergraph-based User Proﬁle in a Recommendation System.

DOI: 10.5220/0005029600270035

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (KEOD-2014), pages 27-35

ISBN: 978-989-758-049-9

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

recommendation case study. The evaluation details

are presented in Section 4. Section 5 concludes the

paper by summarizing the future work.

2 RELATED WORK

The aggregation of partial user proﬁles includes sev-

eral issues such as entity matching, resolution of du-

plicates and conﬂicts, and heterogeneity of the partial

user proﬁles (Orlandi et al., 2012). Furthermore, the

objective of the aggregated user model inﬂuences the

aggregation strategy. In literature, there are diverse

aggregation approaches.

In (Abel et al., 2013), form-based and tag-based

proﬁles are managed separately. The former is a list

of attribute-value pairs whereas the latter is a set of

weighted tags. The aggregation strategy for form-

based proﬁles is unifying sets of attribute-value pairs.

Heterogeneous attribute vocabularies is resolved by

using an alignment function which maps proﬁles to

uniﬁed attribute-value space. However, this align-

ment function may result in duplicate entries in the

ﬁnal user proﬁle. Moreover, when there are conﬂicts

in the aggregated proﬁles, both values are included

in the result. The aggregation of tag-based proﬁles

is accomplished by taking a weighted accumulation

of partial tag-based proﬁles. The authors do not con-

sider aggregating tag-based proﬁles and form-based

proﬁles with each other. In our paper, we do not make

such a distinction. We seamlessly aggregate received

partial user proﬁles by taking their weighted accumu-

lation. We solve heterogeneous vocabulary problem

by using Freebase

In (Orlandi et al., 2012), the aim is to obtain an

interoperable, source-independent, multi-domain user

proﬁle. Therefore, the aggregated user poﬁle is repre-

sented by using popular standard ontologies. During

aggregation the authors address the problem of recur-

ring items and calculating a global weight for them.

To achieve this, they keep track of provenance data

which is the metadata for the user proﬁle item such as

the source of the item and the timestamps. Keeping

track of origins of interest relations enables the re-

calculation of item weights during aggregation of the

partial proﬁles. We also keep track of the provenance

data by storing the knowledge source, the short term

proﬁle date and the exact keyword of the item. We ex-

tend this information each time the item and the user

is bound together.

In (Wischenbart et al., 2012), the aggrega-

tion is handled by semi-automatically extracting

Freebase, https://www.freebase.com/

schema from social web data and integrating the ex-

tracted schemata with existing integration tools. In

(Plumbaum et al., 2011), an aggregation ontology is

proposed to semi-automatically aggregate partial user

proﬁles.

In this paper, the objective of the aggregation is

two-fold: (i) to obtain a user model based on a hyper-

graph which reduces connected data problems such

as recommendation into graph traversal algorithms

and (ii) increasing recommendation accuracy with the

proposed semantic enhancements.

In literature, the semantic enrichment is accom-

plished by disambiguating the concept by linking to

an external vocabulary, using a secondary vocabulary

when the concept could not be linked, enriching the

concept by adding sysnsets, expanding the concept by

retrieving related concepts from the external vocab-

ulary according to a predeﬁned treversal algorithm,

by using friends or like-minded users’ proﬁles as ex-

plained in the survey (Abdel-Hafez and Xu, 2013).

We achive semantic enhancement by using a middle

ontology in front of the external vocabulary and cal-

ibrating the middle ontology concepts according to

system requirements.

Most user modelling and recommendation prob-

lems are connected data problems. Connected data

problems are solved by generating appropriate traver-

sal algorithms which traverse the sub-graph related

to the problem. It is claimed that graph databases

are faster than relational and NoSQL databases when

dealing with connected data (Robinson et al., 2013)

since relational and NoSQL databases lack relation-

ships causing connected data problems to be costly

on these databases. In graph databases, a traversal

query performance depends on the size of the sub-

graph which is going to be traversed. That is, the size

of the whole graph does not effect the traversal per-

formance. Therefore we use a graph database for the

implementation of the hypergraph in this paper.

In (Tiroshi, 2012; Cena et al., 2013), graph based

user models are presented. In (Tiroshi, 2012), con-

cepts are linked to each other by examining an exter-

nal ontology; therefore the nodes could be traversed

in a generic way. In (Cena et al., 2013), horizontal

propogation amongst siblings and vertical propoga-

tion amongst ancestors and decendants are deﬁned.

Our approach is an enhanced version of the former

approach. We connect semantically related concepts

to each other during aggregation process and our pre-

deﬁned node labels and edge types enable deﬁning

more speciﬁc traversal algorithms easily.

KEOD2014-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

3 DATA MODEL AND

AGGREGATION

3.1 Preliminaries

A hypergraph is the generalization of an ordinary

graph by introducing hyperedges which are non-

empty subsets of the vertex set (Gallo et al., 1993).

Vertices of a hypergraph represents the entities to be

modelled such as people and concepts. Hyperedges

represent the high order relations between those enti-

ties.

Besides hypergraphs, there are property graphs

which contain key-value property pairs (Robinson

et al., 2013). In a property graph each node and edge

can have multiple key-value pairs whereas in a hy-

pergraph, an edge can connect more than two nodes.

Every hypergraph can be represented by a property

graph by adding extra key-value pairs to annotate

nodes which are connected by the same hyperedge.

Thus, property graphs are identical to hypergraphs

in terms of representation power. In this paper, we

use property graphs in the implementation, since the

graph database we adopted

supports property graphs.

In the property graph, properties can be indexed

by using a tree like structure. Therefore, a two step

search on graph can be adopted: First the concept is

located in the index structure and then with this short-

cut to the graph, traversal algorithm can be applied.

In graphs, the cost of local read operations is con-

stant, since adjacent vertices and edges are already

connected.

3.2 Hypergraph User Model

We collect short term proﬁles for registered users

from predeﬁned knowledge sources such as facebook,

linkedin for predetermined time periods. Besides, we

allow users to add their interests manually via an in-

terface. In this paper, we focus on constructing a

holistic, multi-domain user model by aggregating the

received short term proﬁles by utilizing the proposed

hypergraph data structure. We use the term partial

proﬁle and short term proﬁle interchangeably in the

paper.

The main components of the user model is sum-

marized in Table 1. In the proposed framework, users,

items and domains are represented with distinct node

types U, I and D. The supported domains are prede-

ﬁned. Freebase commons package

is used as do-

Neo4j, http://www.neo4j.org/

Freebase, https://www.freebase.com/

mains. A domain starter node D

[d]

is created for each

Freebase domain.

In the proposed model, different types of relations

are represented by different edge types. E

bind

is the

edge with label InterestedIn and connects a user u to

an item i to represent that ”u is interested in i”. In

order to model the semantic relations between items,

inner

is used and the label of the edge represents the

nature of the semantic relation. For instance, in Fig-

ure 3 ContributedTo edge is an E

inner

edge which in-

dicates the start node contributes to the end node of

the relation. The item i is connected to its belong-

ing domain d by using E

domain

edge.In the proposed

model, items without any domains are not allowed,

every item must be connected to at least one domain

starter node. The friendship between users is repre-

sented with E

f riend

edges. E

inner

and E

domain

edges

enable content-based recommendations where E

f riend

supports collaborative recommendations.

Table 1: Our Hypergraph User Model.

Notation Description Type

u a user Node

U Set of users Hyperedge

i an item A Node

I Set of items Hyperedge

Domain starter node

[d]

for each Node

domain d

bind

Metadata for user-item Hyperedge

(interest) relation

inner

The semantic relation Hyperedge

between items

The domain bind

domain

between Hyperedge

domain starter node

and items

f riend

Friendship between users Hyperedge

General A sub

(long term) hypergraph

user proﬁle

Deﬁnition. Hypergraph User Proﬁle. The hyper-

graph user proﬁle H

is the aggregated, semantically

enhanced user model for the user u (Eqn.1). It is the

union of the user’s friends whom the user follows or is

followed by (Eqn. 2), the user’s explicit proﬁle which

is the set of user’s declared interested items (Eqn. 3)

and the user’s semantically enhanced proﬁle (Eqn. 4)

The user’s enhanced proﬁle is deﬁned as the set of

items whose shortest path to the user node has at least

min, at most max steps.

UsingHypergraph-basedUserProfileinaRecommendationSystem

(u;min;max) = U

f riends

(u)

∪ U

explicit pro f ile

(u)

∪ U

enhanced pro f ile

(u;min, max)

(1)

friends

(u) = u

f ollows

−−−−→ (u

)

∪ (u

)

f ollows

−−−−→ u

(2)

explicit proﬁle

(u) = u

interestedIn

−−−−−−→ (i)

isInDomain

−−−−−−→ (d)

(3)

enhanced proﬁle

(u;min; max) =

∗min..max

−−−−−→ (i)

isInDomain

−−−−−−→ (d)

(4)

Basically the hypergraph user model consists of

sets of nodes and strongly typed hyperedges. The

proposed hypergraph consists of nodes for domains,

interest items and users; and edges for explicitly

stated interests, semantic relationships between in-

terest items and domain relations of the items. As

an example scenario, assume that there are three

users whose names are GraceKelly, IngridBergman

and TippiHedren. IngridBergman states interest in

three items: Alfred Hitchcock who is a director and

Alfred Hitchcock Presents and The Twilight Zone

which were popular TV shows in 1950s. GraceKelly

expresses interest in the director Alfred Hitchcock

whereas TippiHedren does not declare any interest.

Also these three users are friends. The hypergraph

which models the illustration scenario is in Figure

1; for clarity friendships and domains are eliminated.

The implementation of this hypergraph actually cor-

responds to the property graph shown in Figure 2.

Grace

Ingrid Tippi

Al f redHitch Al f redHitchPres

TwilightZone

HY PER − EDGES

Users

Items

FansO f Al f redHitch

Ingrid

sPro f ile

Figure 1: Illustration Scenario in Hypergraph.

In the hypergraph (Figure 1), the yellow hyper-

edge models the set of users, whereas in the prop-

erty graph (Figure 2) the users are represented with

red nodes. Similarly, the blue hyperedge in the hy-

pergraph is a wrapper for the set of items where the

blue nodes in the property graph are item nodes. The

pink hyperedge in the hypergraph links Ingrid with

her declared interested items. In the property graph,

this hyperedge is modeled by connecting Ingrid to the

items with an edge type InterestedIn. All users are

connected to each other via following mechanism to

represent their friendship. The type of the edge be-

tween users is Follows and the type of edge between

a user and an explicitly declared item is InterestedIn.

Figure 2: Illustration Scenario in Property Graph.

When a new keyword expressing the user’s inter-

est arrives for aggregation, the keyword is located in

the external knowledge base. In this paper, we use

Freebase as the knowledge base and a disambigation

routine which processes the keyword if the keyword

does not match any entity in Freebase. The disam-

biguation routine performs several text pocessing op-

erations. For example it replaces the special charac-

ters with the nearest letters in English alphabet such

as replacing s¸, c¸ by s, c; removes the terms such as

”Fans Of”, ”Quotes” from the keyword; splits the

keyword if it contains characters such as ”&, /”. Free-

base search api returns matching concepts ordered by

score, therefore we used the ﬁrst concept with the

highest score as the matching entity for the keyword.

We deﬁned a domainizer routine to assign the dis-

ambiguated concept to the domains it belongs. In

the proposed model, Freebase domains which corre-

sponds to Freebase commons package is used. For

each domain type, a starter domain node is created at

system initiation. The type information of the con-

cept is retrieved from Freebase. The retrieved type in-

KEOD2014-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

formation not only includes domain knowledge, but

also more speciﬁc type information. For instance,

when the type informtion of Alfred Hitchcock is re-

trieved, types such as Film director, Film producer,

Film writer are also retrieved under the type Film

which is a domain. We exploit those speciﬁc types

to compute the weight of the domain. In other words,

we build a weighted domain structure by accumulat-

ing speciﬁc types under each domain. Afterwards, we

prone the weighted domain structure according to the

predeﬁned domain threshold and relate the concept

with the most frequent domains by using an edge with

type IsInDomain. In Figure 3, the purple nodes rep-

resent the domain starter nodes. There is one starter

node for each domain and all of the items belonging

to that domain is related to that node. This design

facilitates domain-based queries.

The semantic enhancement of a concept

is achieved by retrieving predeﬁned Freebase

Metaschema properties which provides higher order

relations between concepts. Metaschema ontology

consists of 46 properties and constructs another layer

over huge Freebase ontology which has over 3500

properties. Metaschema connects important informa-

tion and eliminates excessively detailed semantics

in Freebase. We further reduced 46 properties to

9 properties by considering their beneﬁts in user

modelling and apply a threshold on the number of

retreived relations. The 9 properties we support

for semantic enhancement include BroaderThan/

NarrowerThan, ContributedTo/ HasContributor,

Created/ CreatedBy, HasGenre/ GenreOf, HasName/

NameOf, HasChild/ HasParent, PractitionerOf/

HasPractitioner, HasSubject/ SubjectOf, Superclas-

sOf/ SubclassOf. Using Freebase over a middle

ontology enables writing domain-independent or

domain-conﬁgured algorithms by using different

thresholds for different domains. For instance, Con-

tributedTo and Created properties reveal important

information for Film and Music domains where

ChildOf property is meaningful in People domain.

The concepts retrieved during semantic enhancement

are related to the key concept with an edge of type

named after the metaschema property linking them.

For instance, in Figure 3, Alfred Hitchcock which is

represented by the blue node at the center is related

to his movies, TV shows and songs with an edge of

type ContributedTo.

3.3 Recommendation Case Study

Various connection-based queries could be answered

by deﬁning traversals on the proposed hypergraph

data structure.

Traversal Example 1. In order to obtain the user

domain-based model for the user u and domain d, the

user is located in the external index system for users

and the user node in the hypergraph is reached with a

short-cut. Eqn. 5 computes user domain-based model

by matching the items which are in domain d and have

a shortest path with the user u with length at most

max.

domain

(u;d; max) = u

∗0..max

−−−−→ (i)

IsInDomain

−−−−−−→ d

(5)

The json output for the query ”Retrieve the do-

main based proﬁle for user GraceKelly for domain

TV.” contains the user’s declared interest Alfred Hith-

cock and the items in her enhanced proﬁle such as

the TV show Alfred Hitchcock Presents and its sev-

eral episodes.

{ "data": [

{ "row": [

"GraceKelly",

"Alfred Hitchcock"

] },

{ "row": [

"GraceKelly",

"Alfred Hitchcock Presents"

] },

{ "row": [

"GraceKelly",

"The Case of Mr. Pelham"

] },

...

] }

To obtain the general user proﬁle, during Traver-

sal Example 1 domain is not included as a parameter

to the traversal function.

Traversal Example 2. In order to discover the users

interested in a domain d, the set of users that have

shortest path with length at most max to d are re-

trieved (Eqn. 6).

domain

(d; max) = d ←− (i)

∗0..max

←−−−− (u)

(6)

Traversal Example 3. To discover users interested

in an item i, the set of users that have shortest path

with length at most max to i are retrieved (Eqn. 7).

item

(i;max) = i

∗0..max

←−−−− (u)

(7)

Traversal Example 4. The ability to discover re-

lated concepts of an item i in other domains as in Eqn.

8 enables answering questions such as ”What are

the ﬁlms about Nasa?” or ”Find biographies about

Mozart.”.

UsingHypergraph-basedUserProfileinaRecommendationSystem

Figure 3: A Sample User Model.

(i;max) = i

IsInDomain

−−−−−−→ (d

)

and i

[∗2..max]

−−−−−→ (d

)

and (otherItem) −→ d

and d

6= d

(8)

Traversal Example 5. In order to calculate a user’s

interest on a concept, shortest path algorithms could

be applied as in Eqn. 9.

interest

(u;i) = shortestPath(u, i)

(9)

Figure 4 shows the interface of the recommendation

system that we implemented based on these traver-

sal algorithms. In the illustration scenario (Figure 2),

GraceKelly declared one interest item: director Alfred

Hitchcock.

The interface is divided into six columns. The

ﬁrst column shows the friendship information, the

second column enables manual addition of an inter-

est item and shows the users declared interests. The

number next to the declared interest is the frequency

of that item and it is incremented by one whenever

the same concept is matched with different keyword-

information source pairs. The list next to the fre-

quency information shows the domains of the item.

The third column exposes the domain aggregation for

the user.

The fourth and ﬁfth columns show the top 15 rec-

ommendations for the user. Random recommenda-

tions part recommends any item which is connected

to the user in the graph via other items or users. De-

tailed recommendations part recommends items that

are connected to the user’s declared items and ranks

the recommendation by checking two factors: the

number of declared items of the user which constitute

a path of length 2 between the user and the recom-

mended item and the accumulated frequency of the

items in that path. For instance, there are two paths

of length 2 between IngridBergman and Mystery item

over the user’s two declared interests: The Twilight

Zone and Alfred Hitchcock Presents. Since both items

are assigned frequency 1, the accumulated frequency

is 2.

Popular recommendations part recommends items

only in popular domains and eliminates other do-

mains. Path length ordering is applied. Far recom-

mendations part recommends items at least three, at

most ﬁve steps away from the user. The sixth column

computes whether the user is interested in the speci-

ﬁed item and lists the users who might be interested

in. For instance, in Figure 4, GraceKelly’s interest for

Marnie, which is a movie directed by Alfred Hitch-

cock, is over declared interest Alfred Hitchcock and

the path length is 2.

KEOD2014-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

Figure 4: Fun Guide Interface.

UsingHypergraph-basedUserProfileinaRecommendationSystem

4 EVALUATION

The dataset is prepared by collecting social web ac-

tivities of 204 users on their Facebook accounts. The

page likes are the main information source. Besides

keywords from shared videos, checkins are also col-

lected. In the populated graph database, there are a

total of 22746 nodes of which 204 are user nodes and

22466 are item nodes. Most popular domains are mu-

sic, awards, people, business, media and ﬁlm and the

most unpopular domains are library, skiing, zoos and

aquariums, bicycles and physics.

The dataset is split into training and testing sets.

80% of each partial proﬁle is located in the train-

ing set and the remaining 20% is left for testing set.

During evaluation, for each keyword on each partial

proﬁle, the keyword is disambiguated by using Free-

base. Keywords which can not be disambiguated are

skipped. After disambiguation, the domains are de-

cided by the domainizer routine by taking the domain

threshold as 4. In other words, each item is linked to

at least 1, at most 4 domains. Then semantic enhance-

ment procedure is applied by taking the threshold 5.

Taking those thresholds higher increase the seman-

tic nature of the constructed model, but the overall

processing time for evaluation also increases. During

evaluation, if the item in the test set is already con-

nected before the interest is declared, this is consid-

ered as a hit. In the evaluation, we considered the ratio

of number of hit items to the number of items of the

user in the short term proﬁle. For 204 users, the aver-

age of hits-to-total items ratio is calculated as 0.61. In

the baseline, the knowledge base usage and enhance-

ment is removed and 40 users of the same dataset is

evaluated likewise. The average hits-to-total ratio for

the baseline is 0.25. The resulting scores show that us-

age of a knowledge base and the enhancement proce-

dure successfully predicts the user’s future interests.

The domain and semantic enhancement thresholds are

kept small to obtain fast evaluation, increasing them

would result in a better hits-to-total ratio in future in-

terest prediction.

Moreover, 20 users of the same dataset is evalu-

ated for cold-start. During evaluation for cold-start,

each user is extracted from the dataset and the hy-

pergraph is populated with the remaining users. Af-

terwards, during aggregation of the user to the pre-

viously populated hypergraph, hits-to-total radio is

calculated. For 20 users, the average of hits-to-total

items ratio is calculated as 0.52. In the baseline, the

average hits-to-total ratio is 0.03. The resulting scores

show that usage of a knowledge base and the enhance-

ment procedure successfully predicts the user’s future

interests in cold-start as well.

In addition, the recommendation study is under

human evaluation currently. The user is able to con-

nect with his/her Facebook account and LinkedIn ac-

count. The system provides Import Facebook and

Import LinkedIn functionality and aggregates the ob-

tained partial proﬁles. The user rates the recommen-

dations provided by the system. The human evalua-

tion system is going to be online for 1 month.

5 CONCLUSIONS

In this paper, we presented a framework for aggregat-

ing partial user proﬁles into a holistic, multi-domain

user model. The main objective of the aggregation

is to obtain a user model data structure which reduces

the connected data problems such as recommendation

into deﬁning graph traversal algorithms.

Graphs naturally support connected data problems

and using property graphs which are equivalent to hy-

pergraphs makes deﬁnition of graph traversal algo-

rithms easier by providing ﬁltering mechanisms such

as node labels and edge types. In other words, it is

possible to write traversal algorithms speciﬁc to a la-

bel or an edge type without traversing irrelevant nodes

or edges in the hypergraph. Another goal of our sys-

tem is to successfully predict user’s future interests.

To achieve this goal, we used an external knowledge

base via a middle ontology and conﬁgured the use of

middle ontology according to user modelling domain.

We only used properties in the middle ontology such

as ContributesTo, Creates, SuperclassOf etc. that are

relevant to user modelling domain.

During evaluation, we showed that the system

could predict future interests of the user with a hit-

to-total ratio of 0.61. If the semantic enhancement

and usage of external knowledge base is eliminated,

the score would be 0.25. As future work, we are go-

ing to categorize users according to social web usage

habits, separate long term and short term user proﬁles

and evaluate the framework against a bigger dataset.

ACKNOWLEDGEMENTS

This work is partially supported by The Scientiﬁc

and Technical Council of Turkey Grant ”TUBITAK

EEEAG-112E111”.

REFERENCES

Abdel-Hafez, A. and Xu, Y. (2013). A survey of user mod-

elling in social media websites. Computer and Infor-

KEOD2014-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

mation Science, 6(4):59–71.

Abel, F., Herder, E., Houben, G.-J., Henze, N., and Krause,

D. (2013). Cross-system user modeling and personal-

ization on the social web. User Modeling and User-

Adapted Interaction, 23(2-3):169–209.

Bu, J., Tan, S., Chen, C., Wang, C., Wu, H., Zhang, L.,

and He, X. (2010). Music recommendation by uniﬁed

hypergraph: Combining social media information and

music content. In Proceedings of the International

Conference on Multimedia, MM ’10, pages 391–400,

New York, NY, USA. ACM.

Cena, F., Likavec, S., and Osborne, F. (2013). Anisotropic

propagation of user interests in ontology-based user

models. Inf. Sci., 250:40–60.

Chen, B., Wang, J., Huang, Q., and Mei, T. (2012). Person-

alized video recommendation through tripartite graph

propagation. In Babaguchi, N., Aizawa, K., Smith,

J. R., Satoh, S., Plagemann, T., Hua, X.-S., and

Yan, R., editors, ACM Multimedia, pages 1133–1136.

ACM.

Gallo, G., Longo, G., and Pallottino, S. (1993). Directed

hypergraphs and applications. Discrete Applied Math-

ematics, 42(2):177–201.

Gauch, S., Speretta, M., Chandramouli, A., and Micarelli,

A. (2007). The adaptive web. In Brusilovsky, P.,

Kobsa, A., and Nejdl, W., editors, The Adaptive Web,

chapter User Proﬁles for Personalized Information

Access, pages 54–89. Springer-Verlag, Berlin, Heidel-

berg.

Ghosh, R. and Dekhil, M. (2008). Mashups for semantic

user proﬁles. In Proceedings of the 17th International

Conference on World Wide Web, WWW ’08, pages

1229–1230, New York, NY, USA. ACM.

Kramar, T., Barla, M., and Bielikova, M. (2013). Person-

alizing search using socially enhanced interest model

built from the stream of user’s activity. J. Web Eng.,

12(1-2):65–92.

Li, L. and Li, T. (2013). News recommendation via hy-

pergraph learning: encapsulation of user behavior and

news content. In Leonardi, S., Panconesi, A., Ferrag-

ina, P., and Gionis, A., editors, WSDM, pages 305–

314. ACM.

Orlandi, F., Breslin, J. G., and Passant, A. (2012). Ag-

gregated, interoperable and multi-domain user proﬁles

for the social web. In Presutti, V. and Pinto, H. S., ed-

itors, I-SEMANTICS, pages 41–48. ACM.

Plumbaum, T., Schulz, K., Kurze, M., and Albayrak, S.

(2011). My personal user interface: A semantic user-

centric approach to manage and share user informa-

tion. In Smith, M. J. and Salvendy, G., editors, HCI

(11), volume 6771 of Lecture Notes in Computer Sci-

ence, pages 585–593. Springer.

Robinson, I., Webber, J., and Eifrem, E. (2013). Graph

Databases. O’Reilly, Beijing.

Tarakci, H. and Cicekli, N. K. (2012a). Ubiquitous fuzzy

user modeling for multi-application environments by

mining socially enhanced online traces. In Masthoff,

J., Mobasher, B., Desmarais, M. C., and Nkambou,

R., editors, UMAP, volume 7379 of Lecture Notes in

Computer Science, pages 387–390. Springer.

Tarakci, H. and Cicekli, N. K. (2012b). Ucasfum: A ubiqui-

tous context-aware semantic fuzzy user modeling sys-

tem. In Filipe, J. and Dietz, J. L. G., editors, KEOD,

pages 278–283. SciTePress.

Tiroshi, A. (2012). Graph based user modeling. In Duarte,

C., Carrio, L., Jorge, J. A., Oviatt, S. L., and Gonalves,

D., editors, IUI, pages 371–374. ACM.

Tiroshi, A., Berkovsky, S., Kaafar, M. A., Chen, T., and

Kuﬂik, T. (2013). Cross social networks interests pre-

dictions based ongraph features. In Proceedings of the

7th ACM Conference on Recommender Systems, Rec-

Sys ’13, pages 319–322, New York, NY, USA. ACM.

Wischenbart, M., Mitsch, S., Kapsammer, E., Kusel, A.,

Prll, B., Retschitzegger, W., Schwinger, W., Schnbck,

J., Wimmer, M., and Lechner, S. (2012). User proﬁle

integration made easy: model-driven extraction and

transformation of social network schemas. In Mille,

A., Gandon, F. L., Misselis, J., Rabinovich, M., and

Staab, S., editors, WWW (Companion Volume), pages

939–948. ACM.

UsingHypergraph-basedUserProfileinaRecommendationSystem