Context-aware Recommender Systems
J.UCS Special Issue
Katrien Verbert
(Departement Computerwetenschappen, K.U.Leuven, Leuven, Belgium
Katrien.verbert@cs.kuleuven.be)
Erik Duval
(Departement Computerwetenschappen, K.U.Leuven, Leuven, Belgium
erik.duval@cs.kuleuven.be)
Stefanie N. Lindstaedt
(Know-Center GmbH, Graz, Austria
slind@know-center.at)
Denis Gillet
(École Polytechnique Féd&ecute;rale de Lausanne, Lausanne, Switzerland
denis.gillet@epfl.ch)
Recommender systems have been researched and deployed extensively
over the last decade in various application areas, including
e-commerce, technology enhanced learning, e-health, adaptive
multimedia and knowledge management. The three approaches of
recommender systems commonly implemented are collaborative filtering,
content-based filtering and hybrid filtering which combines aspects of
both approaches [Balabanovic, 97]. Content-based
recommender systems match content resources to user interests,
typically specified in a user profile. Collaborative recommender
systems recognize commonalities between users on the basis of their
ratings, and generate new recommendations based on inter-user
comparisons. Hybrid recommending approaches combine both content and
user based similarity measures in recommendation algorithms.
Whereas the various approaches have been adopted and validated on a
large scale, not much research has been done to incorporate
contextual information of the user in the recommendation process
[Anand, 07] [Adomavicius,
05]. The context of the user, such as the
task she is working on, time of the day, location and device used,
has a direct impact on the relevance of the recommended items. For
example, if a student is being recommended material to study the
theory of relativity while commuting from the workplace to school
using a smartphone, short, audiovisual material that fits the
screen of the smartphone will be more relevant than a long,
text-only document. Accordingly, the recommender system should rank
short videos higher than long documents. On the other hand, if the
student is reviewing the same subject at night, at home, more
in-depth material, including long texts and formulas, will be more
useful and, consequently, this material should be recommended
first.
This re-ranking of the recommended materials is not possible
if the system does not have any information about the context of
the user. In this perspective, new challenges emerge for capturing
and understanding the context of the user and exploiting such
contextual information for creating intelligent recommendations
adapted to the current, contextual, needs of the user. The
articles contained in this special issue on context-aware
recommender systems present innovative methods, techniques and
systems that allow the measurement, analysis and exploration of
context data and the exploitation of such data to drive
personalized and contextual recommendations.
El Helou,
Salzmann and Gillet present the 3A recommender system that
targets context-aware recommendation in personal learning
environments. The authors give interesting insights into technology
that can be used to extract contextualized user profiles from emerging
information systems. Context is measured and represented by actors,
activity spaces and assets in learning environments as well as
explicit interest parameters such as tags and queries of the user. The
authors propose a contextual and multi-relational ranking mechanism
that adapts a version of Google's PageRank algorithm to the particular
modelling framework, recommending to users not only assets (content),
but also relevant activities and actors to interact with. This
approach is an interesting alternative to existing conceptualizations
of the various actors and components in knowledge sharing
environments, offering contextualized recommendations that consider
all types of relations between actors and components.
Schirru,
Baumann, Memmel and Dengel focus on the automatic
identification of topics a user is interested in and the
identification of thematic groups from which contextualized interest
profiles are built. The approach uses non-negative matrix
factorization (NMF) for resource clustering and has two goals:
delivering recommendations with intra-topic diversity and delivering
targeted context-sensitive recommendations. These context-sensitive
recommendations are based on matching currently visited resources with
available interest topic vectors to enable the generation of
recommendations that can meet the current needs and preferences of the
user.
Stern,
Kaiser, Hofmair, Kraker and Lindstaedt also focus on the
derivation of user context by exploiting the topics a knowledge worker
is currently working on. In addition, the experience level with these
topics is used to tailor recommended content according to the current
needs of the user. A multi-layered associative network is presented
that employs, among others, a domain model to automate the annotation
of resources and fragments of resources. This approach enables to
suggest automatically relevant fragments of resources to the user,
based on her current work context. The article describes in detail the
architecture of a learning material recommender that follows this
approach in the extensive set of APOSDLE tools.
Choi, Lee
and Moon present an alternative approach that is based on the
classification of Web contexts and the application of such a
classification for recommendation purposes. They conducted a thorough
study to identify information quality factors and to relate these
factors to user tasks.
In this study, based on user interviews and survey data analyses,
two Web context groups, "careful" and "casual",
were identified. Results indicate that in a careful Web context, users
have clear target information to seek and want to find relevant and
credible content. In a casual Web context, users want to view new and
popular content even without clear target information. An evaluation
of the proposed algorithm to recognize the current context has been
conducted by gathering and analysing usage data collected by a browser
monitoring module.
Butoianu, Vidal, Verbert, Duval and
Broisin and Niemann, Scheffel, Friedrich, Kirschenmann, Schmitz and
Wolpers also base their work on the monitoring of usage
interactions with tools and resources as a basis to capture context
information. A comparative analysis of approaches to capture such
interactions is presented by Butoianu et al. The authors have compared
several modelling approaches and architectures for managing such data
against several criteria, such as flexibility, extensibility and
scalability. Then, they present a framework that takes into account
the advantages of the presented approaches and that tackles their
limitations.
Finally, Niemann, Scheffel, Friedrich, Kirschenmann,
Schmitz and Wolpers present a new approach for calculating
item-based similarity in order to support collaborative
recommendation. A context-based usage similarity measure is presented
and contrasted with the results of classic content-based item
similarity. The authors then discuss scenarios of how recommendations
may be supported by exploiting the pre- and post-context in which an
item was used.
This special issue is comprised of selected, extended and
peer-reviewed papers presented at the 1st Workshop on Context-aware
Recommendation for Learning at the Second Alpine Rendez-Vous (ARV),
held in Garmisch-Partenkirchen, Germany, on November 30 and
December 1, 2009. The Alpine Rendez-Vous is a series of workshops
that is organized and supported by the STELLAR Network of
Excellence (http://www.stellarnet.eu/) and
aimed at building a
Technology Enhanced Learning researcher capacity on a European
level. The scientific work of STELLAR is organized around three
themes: (1) Connecting learners (2) Orchestrating learning, and (3)
Contextualizing virtual learning environments and instrumentalizing
learning contexts. These themes are intended to be a starting point
for advancing the future of technology enhanced learning. The
contextualization theme addressed in this special issue is an
emerging paradigm for building systems that can anticipate the
needs of learners and act on their behavior. Context-aware
recommender systems are therefore a promising approach for
generating personalized recommendations adapted to the current
needs of the learner and are used for generating suggestions of
relevant learning resources and suitable peer learners who share
similar interests. To that end, these context-ware recommender
systems are also a key enabler towards the "Connecting
Learners" theme. Key research questions include: (a) What
characteristics of users can be exploited to find suitable peer
learners or resources? (b) How to evaluate the efficiency and
effectiveness of context-sensitive recommendations? (c) How to
measure whether learning increased because of the generated
recommendations? and (d) How to deal with key issues of trust,
privacy and security?
These research questions were shaped at the
Alpine Rendez-Vous Workshop on Context-aware Recommendation for
Learning. Thanks to the workshop contributions, links have been
made through STELLAR with several projects on context-aware
recommendation issues, including APOSDLE
(http://www.aposdle.tugraz.at/), MACE
(http://portal.mace-project.eu/) and ROLE
(http://www.role-project.eu/).
We would like to take this
opportunity to thank all the authors who submitted contributions to
this special issue. We also want to thank the external reviewers
for their valuable comments, which contributed significantly to the
high quality of the accepted articles: Nikos Manouselis (Greek
Research and Technology Network, Greece), Hendrik Drachsler (OUNL,
The Netherlands), Martin Wolpers (Fraunhofer FIT, Germany), Jad
Najjar (WU Vienna, Austria), Xavier Ochoa (Escuela Superior
Politecnica del Litoral, Ecuador), Uwe Kirschenmann (Fraunhofer
FIT, Germany), Joris Klerkx (Katholieke Universiteit Leuven,
Belgium), Sandy El Helou (École Polytechnique
Fédérale de Lausanne,
Switzerland), Barbara Kump (Knowledge Media Research Center,
Germany), Viktoria Pammer (Know-Center GmbH, Austria) and Joris
Klerkx (Katholieke Universiteit Leuven, Belgium). This work was
partially financed by the European STELLAR Network of Excellence
(FP7). Katrien Verbert is a Post-doctoral fellow of the Research
Foundation — Flanders (FWO).
References
[Adomavicius, 05] Adomavicius, G., Sankaranarayanan, R., Sen S.,
Tuzhilin, A.: Incorporating contextual information in recommender
systems using a multidimensional approach. ACM Transactions on
Information Systems, 23(1), 103-145, 2005
[Anand, 07] Anand, S. S., Mobasher, B.:
Contextual recommendation. In Lecture Notes In Artificial
Intelligence, volume 4737, 142-160. Springer-Verlag, Berlin,
Heidelberg, 2007
[Balabanovic, 97] Balabanovic, M., Shoham, Y.:
Fab: content-based, collaborative recommendation. Communications of
the ACM, 40 (3), 66-72, 1997
Katrien Verbert
Erik Duval
Stefanie N. Lindstaedt
Denis Gillet
Leuven, Belgium, August 2010
|