This document proposes to use the Friend of a Friend (FOAF) de nition to recommend music depending on user's musical tastes and to lter music-related newsfeeds. One of the goals of the project is to explore music content discovery, based on both user pro ling |FOAF descriptions| and content-based descriptions |extracted from the audio itself.
The World Wide Web has become the host and distribution channel of a broad variety of digital multimedia assets. Although the Internet infrastructure allows simple, straight-forward acquisition, the value of these resources su ers from a lack of powerful content management, retrieval and visualization tools. Music content is no exception: although there is a sizeable amount of text-based information about music (album reviews, artist biographies, etc.) this information is hardly associated to the objects they refer to, that is music pieces. Music is an important vehicle for telling other people something relevant about our personality, history, etc.
In the context of the Semantic Web, there is a clear interest to create a Web of machine-readable homepages describing people, the links between them and the things they create and do. The FOAF (Friend Of A Friend ) project1 provides conventions and a language \to tell" a machine the sort of things a user says about herself in her homepage. FOAF is based on the RDF/XML2 vocabulary. We can foresee that using user's FOAF pro le would allow a system to better understand user musical needs.
The main goal of SIMAC3 project is doing research on semantic descriptors of music contents, in order to use them, by means of a set of prototypes, for providing song collection exploration, retrieval and recommendation services. These services are meant for \home" users, music content producers and distributors and academic users. One special feature is that these descriptions are composed by semantic descriptors. Music will be tagged using a language close to the user's own way of describing its contents |moving the focus from low-level to higher-level (i.e. semantic) descriptions.
Recommender Systems are software applications whose purpose is to deliver information to people that \needs" it. Put this way, one cannot tell the di erence between a Recommender System and a Search Engine |both software types share the same purpose: to select objects (or items) from a repository whose features were found to satisfy the querying user's needs.
However, there exist two subtle but meaningful di erences between \Search Engines" and \Recommender Systems". The rst of these di erences lies in the design intention, or better said: the wording of the problem to address when designing the system. Is that \information need" related to solving a contingent situation, or is that need something periodic or steady? The second one is also another design intention di erence which lies in the use of two di erent words to describe the system: does it retrieve information from a relatively static repository of information? Or does it lter objects embedded in an incoming stream of information?
The \Recommender System" term emerged as a logical evolution of the
research of information retrieval (IR) systems. This evolutive main feature was
the emphasis put on the \query" concept de nition and representation.
Recommender Systems were initially thought as information ltering systems, whose
technological framework baseline stemmed from information retrieval systems
[
There are two main approaches to recommend items to users: collaborative ltering and content-based ltering. Next section explains the di erences between both approximations. 2.1
Collaborative ltering consists of making use of feedback from users to improve
the quality of material presented to users. Obtaining feedback can be explicit
or implicit. Explicit feedback comes in the form of user ratings or annotations,
whereas implicit feedback can be extracted from user's habits. One of the main
caveats of this approach is the fact that the only way to recommend brand new
items is that some user has to previously rate or review that item. are some
examples that succeed based on this approach. For instance, Amazon is a good
illustration system [
On the other hand, content-based ltering tries to extract useful information |from the items of the user's collection| that could be useful to represent user's needs. This approach solves the limitation of collaborative ltering as it can recommend new items (even before not knowing anything from that item), by comparing the actual set of user's items and calculating the distance with some sort of similariy measure. In the music eld, to extract musical semantics from the raw audio and computing similarities between music pieces is a challenging one. Traditional music similarity measures use low-level |mainly timbre-based| features. We belive that adding cultural metadata terms to such a similarity measure can help to get better results. 2.2
The main goal of a music recommendation system is to propose interesting and
unknown music artists (and their available tracks |if possible|) to the
enduser, based on her musical taste. But musical taste and music preferences are
a ected by several factors, even demographic and personality traits. Then,
combining music preferences and personal aspects |such as: age, gender, origin,
occupation, musical education, etc.| could improve music recommendations
[
Moreover, a music recommendation system should be able to get new music dynamically, as it should recommend new items to the user once in a while. In this sense, there is a lot of free available (in terms of licensing) music on Internet, performed by \unknown" artists that can suit perfectly for new recommendations. Nowadays, music websites are noticing the user about new releases or artist's related news, mostly in the form of RSS feeds. iTunes Music Store4 o ers the possibility to subscribe to its New Music Tuesdays system, via email. This service issues one email message every week with exclusives, live session recordings, remixes, celebrity playlists, and unreleased tracks from their artists. iTunes provides, as well, an RSS (version 2.0) feed generator5, with an hourly updated period, that publishes new releases as they are made available. A music recommendation system should take advantage of these publishing services, as well as integrate it into the system, to lter and recommend new music to the user.
Most of the current music recommenders are based on collaborative ltering approach, or an hybrid version including clustering and users' communities Examples of such systems are: Audioscrobbler6, iRate7, Goombah Emergent Music8 and inDiscover9. The basic idea of a music recommender system based on collaborative ltering is to keep track of which artists a user listens |through WinAmp or XMMS plugins|, to in order to nding other users with similar tastes and, nally, recommending similar artists to the user, according on these similar listeners' taste. But, digital music collections can be huge (thousands of les), and very heterogeneous. Thus, this approach to recommend music can generate some \silly" (or obvious) answers.
The main goal of our prototype system is to recommend, to discover and to explore music content; based on both user pro ling |via FOAF descriptions| and content-based descriptions |extracted from the audio itself.
5 http://phobos.apple.com/WebObjects/MZSearch.woa/wo/0.1 6 http://www.audioscrobbler.com 7 http://irate.sourceforge.net 8 http://goombah.emergentmusic.com/ 9 http://www.indiscover.net/
The system is composed by two main components. The rst component is the music recommender, while the second is the (music related) newsfeeds ltering. Both components are based on user's FOAF pro le (example 3.1 shows a possible input le10). Example 3.1 Next sections explains each component of the system.
Music recommendations are done through the following steps: 1. Get interests from user's FOAF pro le 2. Detect artists and bands 3. Access to Music repository and select related artists, from artists encountered in the user's FOAF pro le 4. Rate results by relevance 10 A real example extracted from http://www.livejournal.com, only changing user's name
The prototype reads an input FOAF pro le |that is, an RDF le|, and extracts user's interests. Then queries to a music repository to detect whether the interest is a music artist (or a band) and selects similar artists to the ones found. To get artists' similarities, a focused web crawled has been implemented to look for relationships between artists (such as: related with, in uenced by, followers of, etc.). This web crawler has gathered information from several music portals, such as: allmusic.com, mp3.com and msn.music.com, as well as some sites that contains information (and audio) from \unknown" artists: magnatune.org, garageband.com and vitaminic.com. All these information has been stored into our music repository.
Moreover, a music similarity distance is used to recommend tracks that are similar to tracks composed or played by artists found in the FOAF pro le. Tracks are ltered to the user depending on automatically extracted music content descriptions. These descriptions are composed by a certain number of quanti cable measures taken directly from the track samples ([4], [5]). Currently, we are considering (i) ryhthm descriptors: tempo (beats per minute), meter (binary or ternary) and danceability, (ii) tonal descriptors: tonality, mode and tonality strength, and (iii) timbre descriptors: global loudness and several low-level timbre descriptors. These audio descriptions are evaluated according to a preference model derived from the analysis of users' listening patterns. Euclidean distance |from the instances of the user's model and the potential recommended tracks| is computed to determine whether an incoming track is likely to be listened by the user.
Based on the FOAF example (see example 3.1), the prototype detects the following artists from the user's pro le: Dogs d'Amour, Social Distortion, The Mis ts and The Pogues. Starting from these artists, the system searches for similar artists and artists in uenced by them, and scores them in terms of counting artist occurrences. If there are any tracks in the music repository from artists in the FOAF pro le, it computes the similarity and gets the most signi cant similar tracks from other artists. Figure 2 shows the ouput recommended artists.
To our knowledge, nowadays it does not exist any system that recommends items to a user, based on her FOAF pro le. There is the FilmTrust system11 which is a part of a research study to understand how social preferences might help web sites present information to users in a more useful way. The system collects user reviews and ratings about movies, and holds it into the user's FOAF pro le. Although it has not yet implemented a recommendation system, it includes a rating algorithm for lms based on a trust-based algorithm [6]. 3.2
Filtering news based on user's pro le is another issue related with recommender systems. This kind of system is designed to lter mail, messages from mailing lists, Internet News articles, newswire stories, etc.
In our system, the music related news ltering component queries a newsfeeds system that lters news regarding to related artists found in user's FOAF pro le. To do so, this component permits to communicate with the PubSub server12, via the Jabber protocol, and creates an RSS feed with a given query |that is the user musical preferences found in the FOAF le. PubSub is a matching service that instantly noti es a user when new content is created that matches user's subscription. PubSub reads over over 8 million weblogs, more than 50,000 internet newsgroups and all SEC (EDGAR) lings. Jabber13 is an open secure protocol, an ad-free alternative to consumer instant messaging services like ICQ, MSN, and Yahoo. Jabber makes use of XML protocols that enable any two entities on the Internet to exchange messages, presence, and other structured information in (close to) real time. 11 http://trust.mindswap.org/FilmTrust 12 http://www.pubsub.com 13 http://www.jabber.org
Once the subscription with PubSub.com has been created, it is possible to visualize all the music related news for a given user. Each news item has a bar score that shows how much it is related with user's musical interests. Scoring is done using the TF/IDF ranking algorithm [7]. TF/IDF ranks documents by counting the number of ocurrences of user's term query into each document. 4
The prototype is completely written in PHP. It uses several PHP modules to implement each subsystem: { FOAF: To process RDF les it uses the FOAF class from the Pear framework (PHP Extension and Application Repository)14. This FOAF class makes 14 http://pear.php.net/ use of the RAP module15, thus is easy to access to the document using the RDQL language. { RSS: Newsfeed processing is done using MagpieRSS16 module. MagpieRSS is an RSS and Atom parser for PHP. { Jabber protocol: To interact with PubSub.com server a PHP class has been implemented. This class allows to connect to the server, to authenticate, and to create, retrieve and delete a subscription. This class makes use of class.jabber.php17 module developed by Nathan Fritz.
Access to the prototype is available at: http://www.semanticaudio.org/foa nthe-music 5
We have proposed a system that recommends music and lters music related news based on a given user's pro le. A system based on FOAF pro les allows to \understand" a user in two complementary ways; psychological factors | personality, demographic, socio-economics, situation| and explicit musical preferences. This system, then, is able to lter and to contextualize users' queries.
In the music eld context, we expect that using news ltering about music new releases, artists' interviews, album reviews, etc. can improve a recommendation system in a dynamic way. Finally, this approach opens a wide range of possible usages and applications, such as notifying a user the forthcoming gigs by an artist |playing close to user's location| whose music is similar to user's musical taste. 6
This work is partially funded by SIMAC IST-FP6-507142 European project. 15 http://www.wiwiss.fu-berlin.de/suhl/bizer/rdfapi 16 http://magpierss.sourceforge.net/ 17 http://cjphp.net int.net/ 4. Gouyon, F. Dixon, S. A Review of Automatic Rhythm Description Systems, Computer Music Journal, Vol. 29, Issue 1 - Spring 2005 5. Gomez, E. Herrera, P. Estimating The Tonality Of Polyphonic Audio Files: Cognitive Versus Machine Learning Modelling Strategies, Proceedings of the 5th International ISMIR 2004 Conference, October 2004. Barcelona, Spain. 6. Golbeck, Jennifer, Bijan Parsia: "Trust Network-Based Filtering of Aggregated Claims", to appear in the International Journal of Metadata, Semantics, and Ontologies, 2005. 7. Ricardo Baeza-Yates and Berthier Ribeiro-Neto,: ACM Press, Modern Information Retrieval, Addison-Wesley, 1999.