AN EVENT DISTRIBUTION PLATFORM FOR

RECOMMENDING CULTURAL ACTIVITIES

Toon De Pessemier, Sam Coppens, Erik Mannens, Simon Dooms, Luc Martens

IBBT, Ghent University, Gaston Crommenlaan 8, Ghent, Belgium

Kristof Geebelen

IBBT, K.U. Leuven, Celestijnenlaan 200A, Leuven, Belgium

Keywords:

Event recommendation, Information ﬁltering, Content-based recommendation, Service-oriented architecture.

Abstract:

Today, people have limited leisure time which they want to ﬁll in according to their interests. At the same time,

cultural organisations offer an enormous amount of activities via their websites. This scarcity of time and the

abundance of cultural events reinforce the necessity of recommender systems that assist end-users in discov-

ering events which they are likely to enjoy. However, traditional recommender systems can not cope with

event-speciﬁc restrictions such as the availability, time and location of cultural activities. Moreover, aggre-

gating the events, collecting consistent metadata, and enriching these metadata with cross-domain knowledge

pose additional challenges for the conventional distribution and recommender systems. In this paper, we show

how personalised recommendation, content-based ﬁltering, and distribution of events can be enabled by the

enrichment of events metadata via open linked data sets available on the web of data. For consistency across

several events providers, we propose an event model using an RDF/OWL representation of the EventsML-G2

standard. Integrating these various functionalities as an extendable bus architecture provides an open, user-

friendly event distribution platform that offers the end-user a tool to access useful event information that goes

beyond basic information retrieval.

1 INTRODUCTION

Nowadays, people tend to consult interactive and dy-

namic information from the Internet, rather than static

hard-copy press. This shift applies to the cultural

scene as well: cultural organisations provide detailed

websites which may help to decide how to spend

leisure time optimally. Furthermore, the role of these

cultural websites is evolving from merely information

provider to a guide that puts structure and emphasis

on the demand and supply of cultural events and their

accompanying assets. To deal with the vast and com-

plex structure of cultural events and leisure activities,

we believe that an event model should be used as the

fundamental basis in organizing and accessing cul-

tural event information. To handle the overload of in-

formation, a recommender system based on this event

model is essential to assist users in ﬁnding relevant

activities.

Our ideas are demonstrated in the context of the

end-to-end event distribution platform CUPID

which is implemented for the Flemish cultural scene.

Using multiple information sources, the events are

aggregated and enriched with additional knowledge

from large linked data sets. The metadata of these

events, described by the EventsML-G2 standard, to-

gether with the enriching multimedia assets enable

content-based ﬁltering, recommendation and distribu-

tion of the events.

The remainder of this paper is organized as fol-

lows: Section 2 elaborates on aggregating the events

from various decentralized parties and the enrich-

ment based on available knowledge sources. Sec-

tion 3 describes the utilized EventsML-G2 (Inter-

national Press Telecommunications Council, 2009)

standard completed with its conceptualisation in an

OWL (McGuinness and van Harmelen, 2004) ontol-

ogy as a unifying (meta)data model. This event de-

https://projects.ibbt.be/cupid

231

De Pessemier T., Coppens S., Mannens E., Dooms S., Martens L. and Geebelen K..

AN EVENT DISTRIBUTION PLATFORM FOR RECOMMENDING CULTURAL ACTIVITIES.

DOI: 10.5220/0003193702310236

In Proceedings of the 7th International Conference on Web Information Systems and Technologies (WEBIST-2011), pages 231-236

ISBN: 978-989-8425-51-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

scription is used in the recommender module, which

is discussed in Section 4, to generate personal sug-

gestions. In Section 5 we present the workﬂow-based

bus-infrastructure that is employed to integrate the

different functionalities to a continuous ﬂow of cul-

tural events. Finally, we offer a brief conclusion on

our research work in Section 6.

2 AGGREGATION &

ENRICHMENT

In a ﬁrst phase, event aggregation is performed; i.e.

the retrieval of primary information of cultural activi-

ties from different event organisers. This primary in-

formation is deﬁned as the minimal information that

is necessary to identify and locate the event. In a next

phase, during content enrichment, primary informa-

tion on cultural activities can be extended with more

substantive and facilitating information. Examples of

facilitating information are ‘how the weather will be’

or ‘how to get to the event with public transport’. Sub-

stantive information may include background infor-

mation on the artist, press articles, images, etc.

A lot of event information is distributed by various

content providers all over the Internet. (The Flem-

ish content providers who cooperated in this project

are CultuurNet Vlaanderen and Vooruit). They pub-

lish their events on blogs, web sites, calendars, so-

cial networks, etc. using different formats like sim-

ple HTML

, iCalendar

or RSS

. As a result, the

main challenge of content aggregation in the context

of events is to provide a framework that supports the

aggregation of data from heterogeneous data sources

which use different standards to publish their events.

To deal with these different incoming data for-

mats, a transformation is required to convert the event

data from their original format to the internal event

ontology, discussed in Section 3. For this purpose

we use an XSL transformations engine. XSLT

is a

declarative, XML-based language used for the trans-

formation of XML documents into other XML docu-

ments. To support the aggregation of a new data for-

mat in the future, the XSL style sheet has to be ex-

tended with the mapping of this new format to the in-

ternal event ontology. A double detection mechanism

eliminates redundant data: events with same date and

the same location as existing events are considered as

duplicate and will not be inserted into the database.

http://tools.ietf.org/html/rfc2854

http://tools.ietf.org/html/rfc5545

http://www.rss-speciﬁcations.com

http://www.w3.org/TR/xsltis

Enriching the events is realized by ﬁrst extract-

ing all the concepts out of the event description us-

ing linguistic processing. For this, we utilize Open-

Calais and i.Know’s Information Forensics

service.

Afterwards, the extracted named entities are used for

querying other data sets and incorporating the re-

trieved information. The extracted locations are em-

ployed for querying Toerisme Vlaanderen, GeoNames

and DBpedia

. Toerisme Vlaanderen is a Flemish

tourism information point with a lot of information

on Flemish cities and regions, restaurants, hotels, and

events. The extracted organisations and persons are

utilized for querying the DBpedia data set and Bib-

Net’s data set. BibNet

is a network of Flemish public

libraries, which disseminates bibliographic informa-

tion. The retrieved links are stored and offered as

additional information sources to the end-user. By

enriching the event information, the end-users get a

lot of extra data to make an informed choice. To-

gether with the recommendations, this will help them

to evaluate and select the events which match with

their personal preferences. Another advantage of this

enrichment process is the semantic linking of event

information to other data sets which incorporates the

event information into the ‘Web of Data’.

3 EVENT MODEL

To acquire a standardised procedure for the dissem-

ination of events and to enable content-based rec-

ommendation techniques, a semantic metadata model

for storing and exchanging event information is re-

quired. To represent the events as structured data, var-

ious standards exist. The most commonly used stan-

dard for describing events is iCalendar. This standard,

which describes events for personal management pur-

poses, deﬁnes an event as anything with a scheduled

amount of time on a calendar. Although, the iCalen-

dar format is able to describe a socio-cultural event, it

can not express relations between the events. For in-

stance, a festival can consist of many smaller events,

i.e. concerts or music performances, which have to be

related.

In the Linked Open Data community, the most

used event model is the Music Ontology Events

Model (Centre for Digital Music - University of

London, 2007). AudioScrobbler (LinkingOpenData

(W3C SWEO Community Project) - Centre for Dig-

ital Music, 2007) of Last.FM utilizes this model to

http://www.opencalais.com, http://www.iknow.be

http://www.toerismevlaanderen.be,

http://www.geonames.org, http://dbpedia.org

http://www.bibnet.be

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

232

describe events which are then linked to a user pro-

ﬁle, for which FOAF

is used. This way, an event is

modelled whenever a song is played. Originally, this

model was intended to describe musical events, but

due to its simplicity and usability, it has been proven

useful in a wide range of contexts. This model de-

scribes an event as anything that has a spatial and

temporal dimension. Such an event is described by

its participating agents, its passive factors inﬂuenc-

ing the event, its products as a consequence of the

event and a location in time as well as space. In addi-

tion, this model allows describing relations between

events. However, its simplicity is also a disadvantage

since the model lacks some advanced features, like

pricing information, more detailed relations between

events, minimum age for participation, etc., which are

essential for describing events in our context.

Finally, the format that we adopted for mod-

elling events is the EventsML-G2 (International Press

Telecommunications Council, 2009) standard. This

is a standard of the International Press Telecommuni-

cations Council

(IPTC) for conveying event infor-

mation in a news industry environment. EventsML-

G2 is intended for receiving, storing and exchang-

ing event information from organisers as well as pub-

lishing event information by news providers. This

model delivers the right context for our event descrip-

tions since it allows describing events in different lan-

guages, together with their relations, the pricing in-

formation, the minimum age, etc.

Currently, the EventsML-G2 standard is described

as an XML (Bray et al., 2006) schema. In order to ap-

ply semantic web techniques, we developed an OWL

(McGuinness and van Harmelen, 2004) ontology of

this standard. An alternative to make the EventsML-

G2 schema usable in the Semantic Web Stack, is de-

scribing the schema in RDFS (Brickley, 2004). How-

ever, since OWL is a richer language for develop-

ing models and permits to mix RDFS and OWL con-

structs, we opted for describing the EventsML-G2 se-

mantic ontology by using OWL. This OWL ontology

is published online on the website of the Multimedia

Lab research group of Ghent University

By describing all the aggregated events using this

ontology, a common format is created for exposing

the event information. This common format acts as

a unifying layer, relating all the information com-

ing from different data providers. By providing a

semantic ontology for these events, content-based

recommendation algorithms are able to analyse the

http://xmlns.com/foaf/spec/

http://www.iptc.org

http://multimedialab.elis.ugent.be/ontologies/

EventsML-G2/v1.0/

events in detail, and additionally, semantic web tech-

niques can be used. Moreover, this semantic ontol-

ogy allows complex queries for events by a SPARQL

(Prud’hommeaux and Seaborne, 2007) endpoint and

the enrichment of event descriptions with information

from other data providers as described in Section 2.

EventsML-G2 has two manners for conveying

event information: as a conceptItem or as a knowl-

edgeItem. A conceptItem is aimed at describing

an event solely. A knowledgeItem is intended for

bundling a set of events which are managed as a

whole. Given our context, i.e. publishing and recom-

mending events from event organisers, we utilize the

conceptItem to model the events. For interoperability

issues, we modelled agents involved in an event, e.g.

organisers or participants, using the FOAF ontology.

This allows us to incorporate more easily information

about the agents from other data sets afterwards. This

additional advantage is the main difference between

our developed ontology and the EventsML-G2 speci-

ﬁcation.

The descriptive metadata about the event is cap-

tured by the Concept class, which stores a description

of the event, the title of the event, the event details, re-

lations to other events, and the language of the event

description. The relations to other events can be de-

scribed by owl:sameAs for stipulating that this event

is the same as another event, and by skos:related,

skos:broader or skos:narrower for pinpointing the re-

lations to other events. The EventDetails class con-

tains a more complete event description and is linked

to ten other classes for storing detailed information

about the event, namely: Dates, Location, Contact-

Info, Agent, Language, Subject, OccurStatus, Regis-

tration, ParticipationRequrement, and Media.

4 EVENT RECOMMENDATION

To handle the overload of information, a recom-

mender system is necessary to assist users in ﬁnd-

ing the most relevant events. Traditionally, recom-

mender systems have been categorized into two main

classes: collaborative ﬁltering (CF) and content-based

(CB) methods. CF techniques are based on the hy-

pothesis that a good method to ﬁnd interesting con-

tent is to search for other people who have similar

interests, and then recommend items that those sim-

ilar users like (Breese et al., 1998). Most existing

recommender applications utilize these CF techniques

in making predictions about which items an e-service

user is likely to be interested in (Linden et al., 2003).

Traditional CF techniques, however, can not cope at

all with time speciﬁc items, like events, which typ-

AN EVENT DISTRIBUTION PLATFORM FOR RECOMMENDING CULTURAL ACTIVITIES

233

ically receive their ratings only after they have ﬁn-

ished (Cornelis et al., 2005). Content-based (CB)

algorithms, which consider the internal structure of

items and recommend items similar to those a user

liked in the past, or hybrid algorithms, which are a

combination of CF and CB techniques, can make up

for these timely characteristics of events.

Although the details of various systems differ,

CB recommender systems share in common a need

for a description of the items that may be recom-

mended (Pazzani and Billsus, 2007). Many CB rec-

ommender systems use relatively simple keyword ex-

traction techniques to obtain characteristic properties

from these descriptions. Subsequently, straightfor-

ward retrieval models, such as keyword matching, are

utilized to generated recommendations (Bogers and

van den Bosch, 2007). However, these techniques

are known to have been outperformed by newer tech-

niques that incorporate semantic knowledge struc-

tures, such as ontologies, to provide valuable domain

knowledge for CB recommender systems (Middleton

et al., 2003). Common-sense and domain-speciﬁc

knowledge may be useful to give some meaning to

the content of items, thus helping to generate more

informative features than ‘plain’ attributes.

Knowledge sources, such as structured semantic

information about events, can be brought to bear in

determining similarities among events. The integra-

tion of semantic similarities for items allows the sys-

tem to make inferences based on the underlying rea-

sons for which a user may or may not be interested

in a particular event (Mobasher et al., 2003). Ex-

perimental results in literature demonstrate that the

integration of these ontological knowledge sources

yields signiﬁcant advantages in terms of learning user

proﬁles and recommendation accuracy in contrast to

the classical use of endogenous knowledge (extracted

from the items themselves) (Semeraro et al., 2009).

However, acquiring such knowledge sources or

semantically structured content is not a trivial task.

Moreover, content (and especially event information)

from various sources is mostly annotated using a mul-

titude of different metadata formats, leading to se-

rious comparison difﬁculties for the recommender.

By converting the aggregated events to our proposed

event model, we provide a uniform event structure

that can be utilized for semantic analysis. By enrich-

ing the event description based on a set of knowledge

sources, we generate a memory of semantic compe-

tencies that can be exploited to reason about the con-

tent as well as to support the user proﬁling and recom-

mendation process. Moreover, based on our service-

oriented bus-architecture as described in Section 5,

we can easily integrate any existing (ontology-based)

recommendation algorithm in our event distribution

system.

To efﬁciently calculate the personal suggestions

based on the adopted (ontology-based) algorithm for a

potentially high number of users, we designed the rec-

ommendation framework which is illustrated in Fig-

ure 1. To meet the high requirements of a large rec-

ommendation system, we opted for a highly scalable

approach to store and process user and event metadata

completed with online-generated user interaction be-

haviour (i.e. user feedback on the events). Moreover,

due to the computational burden of most recommen-

dation algorithms, the calculation of personal sugges-

tions has to be realized as a distributed computation

task.

The proposed recommender framework consists

of two webservices and three small, independent ap-

plications working together to collect user behaviour

and metadata content from participating web sites

and to generate high quality recommendations for

their visitors. The webservices act as a communica-

tion channel between the recommendation system and

the other parts of the event distribution platform (i.e.

client websites or other webservices). The Data In-

sertion webservice deals with the processing of new

users, events and feedback on events, which are fed

into the recommender. Based on these data, personal

event recommendations are generated which can be

queried through the Recommendation webservice.

Each of the three applications accomplishes a sin-

gle well-deﬁned task. The ﬁrst application, the Data

Inserter, inserts users, events and user feedback into

the storage system. Secondly, the Calculator pulls

these data from the storage system and generates per-

sonal suggestions based on the implemented recom-

mendation algorithm. Next, these suggestions are

stored in the Recommendation Cache, making them

immediately available for retrieval by the Recommen-

dation webservice. Depending on the work load, sev-

eral instances of this Data Inserter and Calculator

may be created. At regular intervals, the third appli-

cation, i.e. the Scheduler, generates a calculation task

which is split into disjunctive jobs for the Calcula-

tor instances. The results of these calculation jobs are

merged and inserted into the Recommendation Cache

as an update of the personal suggestions.

Because of performance reasons, the different ap-

plications are not in direct contact with each other,

but use a queue service for inter-process communica-

tion. Moreover, the storage system is optimized for

efﬁciently serving two commonly-used data requests:

on the one hand inserting singular user, event or feed-

back information originating from the Data Insertion

webservice, and on the other hand, querying massive

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

234

Scalable Calculation Cloud

Storage

Clients

Webservices

Data Insertion Web Service

Recommendation Web Service

Data Inserter

Calculator

Scheduler

Recommendation

Cache

Client Website

Other Services

Recommendation

algorithm + filters

Figure 1: The recommendation framework.

feedback and event information to feed the instances

of the Calculator application.

Besides the storage and calculation functionality,

the recommendation framework consists of various

contextual post-ﬁlters which take into account the per-

sonal event selection criteria. These ﬁlters operate on

the list of personal event recommendations and can

eliminate the suggestions which do not conform to

the personal event selection criteria. The personal

event selection criteria, which can explicitly be spec-

iﬁed by the end-user, are personal requirements re-

garding the price, language, location, participation re-

quirements or date of the event. E.g., events that are

located too far from the hometown of the user are re-

moved from the suggestions list, since they are mean-

ingless for the end-user. Accordingly, suggestions

which completely fulﬁl all the personal requirements

are favoured during the generation of the recommen-

dation list. Through the proposed event ontology, as

discussed in Section 3, and the enrichment process,

as explained in Section 2 event suggestions can eas-

ily be veriﬁed against the stated event requirements

before distribution.

5 ARCHITECTURE

Figure 2 gives a schematic overview of the infras-

tructure of the complete event distribution platform,

a loosely coupled bus-architecture consisting of 4

components: aggregation, enrichment, recommenda-

tion, and distribution. This architecture supports the

ﬂow of personalized, enriched cultural activities and

connects the various functional components of our

system. An enterprise service bus (ESB) provides

an open, standards-based connectivity infrastructure

for this Service-Oriented Architecture (SOA)

. The

SOA enables ﬂexible connectivity of applications or

resources by representing every application or re-

source as a service with a standardized interface. This

ﬂexibility allows new and existing applications to

http://opengroup.org/projects/soa/

Figure 2: The service bus architecture.

be easily and quickly combined to address changing

business needs.

In the context of the event distribution system,

this means that the different components that com-

pose the general architecture are modelled as sepa-

rate services where the SOA allows them to exchange

data with one another as they participate in business

processes. The main advantage of this SOA is to ob-

tain a loose coupling between the different modules

of the system in which services can be very heteroge-

neous. The implementation of these services is lan-

guage and technology independent; only communi-

cations must be agreed on. The high ﬂexibility sup-

ported by this technology allows an easy integration

of new sources (e.g. aggregation sources or distribu-

tion targets) and remove or modify existing sources

without interrupting the working system. In addition,

using a standardized technology provides also the ad-

vantage of reusable existing middleware services. Ex-

isting implementations contain many pre-built com-

ponents which can be reused with a minimal effort.

Existing components are available in the domain of

logic and orchestration, databases and data manipu-

lation, basic interfacing, security, logging, etc. Al-

though an advanced SOA setup requires some addi-

tional start-up costs and a slight increase of complex-

ity, it provides the freedom to expand and modify the

active setup at any time while requiring only a small

amount of effort.

The ESB technology we used for this project is

the Open Enterprise Service Bus

. OpenESB is an

open-source project with the aim of building an En-

terprise Service Bus that provides a ﬂexible and ex-

tensible platform on which to build SOA and applica-

tion integration solutions. Pulling event information

from the web is triggered by a scheduler. On prede-

ﬁned times, the scheduler sends a message to the input

services to check the input sources for new content.

Orchestration between services is handled by a work-

ﬂow engine, which is integrated in the service bus ar-

chitecture and supports the execution of WS-BPEL

processes (Business Process Execution Language for

https://open-esb.dev.java.net/

AN EVENT DISTRIBUTION PLATFORM FOR RECOMMENDING CULTURAL ACTIVITIES

235

Web Services

). WS-BPEL is built on top of XML

speciﬁcations and provides a language for the spec-

iﬁcation of business processes and business interac-

tion protocols. An executable WS-BPEL process is

deﬁned by a control ﬂow that consists of a combi-

nation of basic and structured activities. WS-BPEL

(and SOA in general) supports the Simple Object Ac-

cess Protocol (SOAP), which is a speciﬁcation for ex-

changing structured information between services.

6 CONCLUSIONS

Online event-data from various providers may emerge

as a dynamic mix of many different types of re-

sources and facilities. This multiplicity of technolo-

gies and metadata descriptions requires a uniform

event model and dynamic event distribution archi-

tecture. In this paper, we proposed an event model

based on the EventsML-G2 standard and a service-

oriented bus architecture for event recommender and

distribution systems. Aggregation, enrichment, rec-

ommendation and distribution are supported through

services, which can be created and modiﬁed dy-

namically to suit the current needs. This physical

loose-coupling provides scalability advantages such

as high-availability, fault-tolerance, and load balanc-

ing. The combination of the semantic event model

and service-oriented architecture allows the applica-

tion of an exchangeable content-based recommenda-

tion algorithm. Enriching the event metadata enables

the incorporation of semantic knowledge structures

into the recommendation algorithm.

ACKNOWLEDGEMENTS

The research activities that have been described in this

paper were funded by Ghent University, K.U. Leuven,

VUB, VRT-medialab, Interdisciplinary Institute for

Broadband Technology (IBBT) through the CUPID

project (50% co-funded by industrial partners), the In-

stitute for the Promotion of Innovation by Science and

Technology in Flanders (IWT), the Fund for Scien-

tiﬁc Research-Flanders (FWO-Flanders), and the Eu-

ropean Union.

http://docs.oasis-open.org/wsbpel/

REFERENCES

Bogers, T. and van den Bosch, A. (2007). Comparing and

evaluating information retrieval algorithms for news

recommendation. In RecSys ’07: Proceedings of

the 2007 ACM conference on Recommender systems,

pages 141–144, New York, NY, USA. ACM.

Bray, T., Paoli, J., Sperberg-McQueen, C., Maler, E., and

Yergeau, F., editors (2006). Extensible Markup Lan-

guage (XML) 1.0 (Fourth Edition). W3C Recommen-

dation. World Wide Web Consortium. Available at

http://www.w3.org/TR/2006/REC-xml-20060816/.

Breese, J., Heckerman, D., and Kadie, C. (1998). Empir-

ical Analysis of Predictive Algorithms for Collabora-

tive Filtering. In Proceedings of the 14th Conference

on Uncertainty in Artiﬁcial Intelligence, pages 43–52,

Madison, USA.

Brickley, D., editor (2004). RDF Vocabulary Descrip-

tion Language 1.0: RDF Schema. W3C Recommen-

dation. World Wide Web Consortium. Available at

http://www.w3.org/TR/rdf-schema/.

Centre for Digital Music - University of London

(2007). The Event Ontology. Available at

http://purl.org/NET/c4dm/event.owl.

Cornelis, C., Guo, X., Lu, J., and Zhang, G. (2005). A

Fuzzy Relational Approach to Event Recommenda-

tion. In Proceedings of the 1st Indian International

Conference on Artiﬁcial Intelligence, pages 2231–

2242, Pune, India.

International Press Telecommunications Council (2009).

EventsML-G2 Speciﬁcation – Version 1.1. Available

at http://www.iptc.com/std/EventsML-G2/EventsML-

G2 1.1.zip.

Linden, G., Smith, B., and York, J. (2003). Amazon.com

Recommendations: Item-to-item Collaborative Filter-

ing. IEEE Internet Computing, 7(1):76–80.

LinkingOpenData (W3C SWEO Community Project) -

Centre for Digital Music (2007). Audioscrobbler RDF

Service. Available at http://dbtune.org/last-fm/.

McGuinness, D. and van Harmelen, F., editors (2004). OWL

Web Ontology Language: Overview. W3C Recom-

mendation. World Wide Web Consortium. Available

at http://www.w3.org/TR/owl-features/.

Middleton, S. E., Shadbolt, N., and Roure, D. D. (2003).

Ontology-based recommender systems.

Mobasher, B., Jin, X., and Zhou, Y. (2003). Semanti-

cally enhanced collaborative ﬁltering on the web. In

Proceedings of the 1st European Web Mining Forum

(EWMF2003), pages 57–76.

Pazzani, M. and Billsus, D. (2007). Content-Based Recom-

mendation Systems. Springer.

Prud’hommeaux, E. and Seaborne, A., editors (2007).

SPARQL Query Language for RDF. W3C Recom-

mendation. World Wide Web Consortium. Available

at http://www.w3.org/TR/rdf-sparql-query/.

Semeraro, G., Lops, P., Basile, P., and de Gemmis, M.

(2009). Knowledge infusion into content-based rec-

ommender systems. In RecSys ’09: Proceedings of

the third ACM conference on Recommender systems,

pages 301–304, New York, NY, USA. ACM.

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

236