Fourth IEEE International Conference on eScience

Mendeley – A Last.fm for Research?

Victor Henning Jan Reichelt

Bauhaus-University of Weimar, Germany University of Cologne, Germany
Mendeley Ltd., London, UK Mendeley Ltd., London, UK
victor.henning@mendeley.com jan.reichelt@mendeley.com

Abstract and musical preferences. The data is gathered as follows:

Last.fm’s “Audioscrobbler” desktop software, after
This paper aims to explore how the principles of a having been installed on a user’s PC, starts tracking a
well-known Web 2.0 service, the world’s largest social user’s music listening behavior. The listening data is sent
music service “Last.fm” (www.last.fm), can be applied to to the Last.fm website, where a profile of the user’s
research, which potential it could have in the world of musical tastes is created. Listening statistics for each
research (e.g. an open and interdisciplinary database, song, album, artist, and genre are aggregated and made
usage-based reputation metrics, and collaborative available online. In this way, Last.fm has created the
filtering) and which challenges such a model would face world’s largest open music database, comprising over 80
in academia. A real-world application of these million songs, accessible by everyone. The user-
principles, “Mendeley” (www.mendeley.com), will be generated data also lays the foundation for
demoed at the IEEE e-Science Conference 2008. personalization, collaborative filtering, and ontological
classifications:
· Users can view timelines and statistics about their
1. Introduction own listening behavior,
· view the most popular tracks for each of their
Ways of how to turn Web 2.0 applications into favorite artists, and most popular artists for their
productive social research tools are currently being favorite genre,
discussed at major academic conferences (e.g. European · receive music recommendations based on the song
Science Open Forum 2008, Science Blogging Conference library already existing on their PC, and
2008, Science in the 21st Century Conference), and both · discover similar tracks/artists for every track/artist in
multi-purpose social software, such as wikis, blogs, and the Last.fm database.
social networks, and more specific services such as
Twitter, Friendfeed, or CiteULike are currently being 3. The model of Last.fm applied to research
used and evaluated by a number researchers and
academics. This paper aims to explore how the principles Last.fm’s service is based on aggregating the users’
of a well-known Web 2.0 service, “Last.fm”, can be existing music libraries, relationships between artists
applied to the domain of academic research. writing songs in different genres, and the users’ music
listening behavior. Similarly, a service for academic
2. How Last.fm works researchers could be based on aggregating scholars’
existing research paper libraries, relationships between
Last.fm (www.last.fm), which bills itself as a “social researchers writing papers in different disciplines, and
music service”, has managed to create the largest the scholars’ paper reading behavior.
ontological classification (and the largest open database) Along these lines, a “Last.fm for research” would be
of music in the world, by aggregating the musical tastes able to display statistics to each individual user about his
of its 20 million users and then data-mining it for similar personal library, to aggregate readership statistics about
musical genres, artists, and songs. The users form a papers, authors, journals, and academic disciplines, and
social network that is not based on pre-existing real- to recommend interesting articles and researchers to the
world relationships; instead, Last.fm’s network emerges user. We envision that such a tool consists of two parts:
around data that describes its users’ listening behavior First, a desktop application which helps researchers

978-0-7695-3535-7/08 $25.00 © 2008 IEEE 327

DOI 10.1109/eScience.2008.128
manage their academic papers and anonymously tracks linguistics, neurophysiology). However, due to space
their reading habits and literature usage. Second, a constraints, he would mostly limit his citations to papers
website where the users can discover aggregated published in psychology journals. Picking up such
statistics, top papers, trends and charts for each patterns, CF recommendations would thus help
discipline, paper recommendations, and introductions to researchers to discover literature that could be of interest
people with similar research interests. Adoption of such a to them, even though it is not found in the citation
service would have a number of advantages for academia network of their existing library. Moreover, CF would
at large, of which we will discuss three important ones. enable researchers to identify people with similar
The creation of an open and interdisciplinary research interests (based on their paper libraries) and
database: Similar to Last.fm’s efforts in the space of thus foster collaboration and academic networking.
music, a tool which aggregates metadata, tags and article Finally, CF would start generating a rich network of
usage of a large number of researchers could lead to an relationships for a paper as soon as it is published, rather
open, interdisciplinary and ontological database of than having to wait months or years to get cited.
research, providing a free and invaluable source of Of course, there are a number of obstacles to be
information to every individual researcher. Working in overcome before such a model can be turned into reality
conjunction with Open Access libraries, this would be in the field of academic research. Arguably the biggest
another cornerstone in building alternatives to expensive obstacle is that a sufficient number of participants is
pay-walled databases. needed to gather reliable usage data. So how could
Usage-based reputation metrics and real-time scholars be convinced of taking part in generating
statistics: Usage-based reputation metrics, which could research paper usage data? In our opinion, the answer is
be described as “Nielsen ratings for science”, would that doing so must confer some type of utility to them
alleviate many of these problems associated with beyond the idea of contributing to a fuzzy “greater good
traditional citation-based reputation metrics. A starting of science”. More specifically, the tool that does the
point for usage-based metrics would be to track the measuring on the researchers’ computers should do this
pervasiveness of research papers, i.e. whether they are only as a secondary purpose, and must have some other
present on the computers of a wide-ranging, distributed primary usage value. Moreover, whereas Last.fm openly
sample of academics. This would be a measure of the displays each user’s listening behavior, privacy is critical
popularity or awareness that a paper – and by association, in the space of research. While it isn’t much of a
its author, publication journal, and topic – is enjoying. A technical issue to hide a researcher’s library and reading
second, more fine-grained usage metric would be the data, a “Last.fm for research” would have to convince
actual time spent reading each research paper (on screen, potential users that it can be trusted with such sensitive
e.g. in Adobe Reader), and the number of repeat readings data, and that no personally identifiable data would ever
per paper. This would be a measure of the intensity with be made public without the researcher’s explicit consent
which the paper (its author, publication journal, topic) is
being examined – did readers only skim through a paper, 4. Mendeley – A Last.fm for research
or did they peruse it in detail? Finally, these metrics
could be augmented with quality ratings and tags that Mendeley is an application of Last.fm’s principles to
help differentiate mere measures of attention from the domain of research. Mendeley Desktop, a free and
explicit quality judgments. cross-platform desktop application, automatically extracts
Collaborative filtering: Usage data would also be the metadata, full-text and cited references from research
basis for developing paper recommendation engines papers to minimize manual data input when setting up a
based on collaborative filtering (CF) principles. It has local research paper database. It then enables researchers
been argued that citations already act as reading to manage, tag, full-text search, cite in Word and LaTeX,
recommendations, and recommendation models based on and share research papers, thus providing researchers
co-citations or citation networks (“localized graph with usage value independent of any network effects.
search”) already exist. CF recommendations, however, The companion website, Mendeley Web, can be used for
have additional potential. First, they could increase the backing up research papers, creating a public research
interdisciplinarity of research because they may be better profile, and connecting to like-minded researchers.
in uncovering parallels between academic disciplines Mendeley Web already displays the pervasiveness of
than citation networks. For example, when doing research papers, authors, journals and tags as measured
research on the psychology of emotion, a psychologist by Mendeley Desktop. Reading time and quality rating
would typically also have read papers on emotion metrics as well as CF recommendation mechanisms will
published in adjacent fields (philosophy, literature, be implemented soon.

328