Current music recommender systems rely on techniques like collaborative ltering on user-provided information in order to generate relevant recommendations based upon users' music collections or listening habits. In this paper, we examine whether better recommendations can be obtained by taking into account the music preferences of the user's social contacts. We assume that music is naturally di used through the social network of its listeners, and that we can propagate automatic recommendations in the same way through the network. In order to test this statement, we developed a music recommender application called Starnet on a Social Networking Service. It generated recommendations based either on positive ratings of friends (social recommendations), positive ratings of others in the network (nonsocial recommendations), or not based on ratings (random recommendations). The user responses to each type of recommendation indicate that social recommendations are better than non-social recommendations, which are in turn better than random recommendations. Likewise, the discovery of novel and relevant music is more likely via social recommendations than non-social. Social shu e recommendations enable people to discover music through a serendipitous process powered by human relationships and tastes, exploiting the user's social network to share cultural experiences.
WOMRAD 2011 2nd Workshop on Music Recommendation and Discovery, colocated with ACM RecSys 2011 (Chicago, US) Copyright c . This is an open-access article distributed under the terms of the Creative Commons Attribution License 3.0 Unported, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
An interesting mechanism of music discovery occurs when
family and friends recommend each other music that they
discover. The emergence of Social Networking Services (SNS)
and Web Music Communities (WMC) provide us the
opportunity to develop music recommender applications to
support this mechanism. SNS are becoming increasingly
popular means for people to socialise online. Music is playing
a similar role on these platforms as in real life social
networks. It is shared, discussed, recommended and discovered
with social contacts. It is noteworthy that music
information seeking behaviour has been indicated as `highly social'
[
In this work, we model that music discoveries take place 1http://www.last.fm 2http://www.pandora.com 3http://www.apple.com/itunes/ping 4http://www.facebook.com via natural di usion of music through social networks or randomly. An experiment was conducted to reproduce this process so that we could analyse how people respond to the recommendations. The recommendations arising from such processes are either randomly picked from the pool of tracks of the data set or collaboratively from the tracks recommended by other people on the SNS. A successful music discovery occurs when the user of the application likes a track that s/he has never heard before.
This paper has been divided into 7 sections; section 1 is the introduction, section 2 elaborates the rationale behind this research experiment. In section 3, we discuss the methodology. In section 4, the results are presented. In sections 5, 6 and 7, limitations are discussed, the research work is concluded, and future work is proposed respectively.
Research on nding new music shows that music discovery
often occurs with the personal acquaintances playing music
to the respondents, and that social networks continue to
play a vital role in music discovery in the digital age [
The aim of the experiment is to test that discoveries are di used through the social network. The problem is treated as a recommendation problem de ned as follows: given a pool of items, select an item that the subject has not heard and that is relevant to her/him. A collaborative recommendation makes use of items rated previously by other subjects to choose the item to be recommended. If collaborative recommendations based on ratings by people from the social network of the subject lead to more successful recommendations than ratings from people not in the subject's social network, then there is an indication that social recommendations are more appropriate for collaborative recommendations. 3.1
The idea of the social shu e is to recommend tracks (see
Section 3.2 for more details) and di use discoveries through
the social network. Figure 1 shows an example of this
process when a recommendation is posted by a user in her/his
SNS. In this case, Joe gets a random recommendation, gives
it a 4 star rating, the track is then di used to his social
network. If a friend of Joe gives a high rating to the same track,
it will be di used to her network as well (in this case, the
example of Alice). If Joe's friend does not enjoy the track
and gives it 2 or less rating then the social di usion stops
and here Aleks' friends will not get this recommendation.
Each time a track is recommended, the subject of the
experiment rates the track on a 0 to 5 Likert scale; a psychometric
scale introduced by Rensis Likert [
In order to have a representative pool of available tracks for random selection, the dataset was built from Last.fm containing a million tracks fetched through their API. We explored a subset of the Last.fm of about 300,000 people and their friends. The tag pro le of each user was fetched, giving us the list of tags they used to organize tracks, yielding a set of more than 300,000 unique tags. For each tag the API allows to fetch the top 50 most popular tracks tagged with this tag. Only the set of tracks is relevant for this study. On Facebook, we started by fetching the network of the rst author, his friends and friends of friends interested in participating in this experiment. Many people share music using Youtube6 by posting links on Facebook. The advantage of using music videos from Youtube is that the full track can be played (if the video is available), which is not the case for public users of Last.fm for instance which limits the playback to 30 seconds. The limitation pertaining to Youtube videos is that sometimes these are blocked in various countries and 5http://www.facebook.com/press/info.php?statistics 6http://www.youtube.com removed for copyright violations. The application detects when such errors occur and does not recommend that video afterwards. On the Youtube API, videos were searched for each track with the artist name and track title. We selected the most popular video for each track and restricted to fetch the videos tagged as \Music" only. Using this process about a quarter of the tracks from Last.fm had a video on Youtube, totaling around 252,000 tracks. 3.1.2
Starnet is a Facebook application fed by ratings on random selections. A positive rating spawns di usion through the user's social network, i.e. the shu e recommendation becomes social. The interface (Figure 2) consists of the current track description (title and artist name), the music video associated, a tag cloud of the user's pro le and a rating form consisting of 5 stars. Also, it has \next" button to play the next track and a \bail" button which sets the stars to 0 and plays the next track. The user is asked to indicate whether the track is already known, in order to learn whether it is a music discovery. We assume it is a reasonable estimate for true unknown tracks.
The subjects are the people who used the Starnet application. A total of 68 subjects participated in this experiment and allowed access to their social pro les. Each subject acted as his/her own control group by getting and rating random recommendations. In about 4 months (from the 29th June 2010 to the 18th October 2010), 31 subjects made 4966 ratings using the Starnet Application. Participation of other 37 subjects was insigni cant in terms of ratings. 3.2
The recommender system produced the following types of recommendations:
Random recommendations: The random selection is a query which selects tracks that have not been rated by the subject and orders them randomly.
Collaborative recommendations: Selects a track randomly from the set of tracks that have been rated with a rating above average (i.e. greater than 2 stars) and not yet rated by the subject. The collaborative recommendation can be social or non-social.
Social recommendations: The social recommender selects a track randomly, from the tracks that have been rated by friends of the subject, with a rating superior to 2 stars.
Non-social recommendations: The non-social recommender selects a track randomly, from the tracks that have been rated by people who are not friends of the subject, with a rating superior to 2 stars.
The likelihood for selecting random recommendation, social recommendation and non-social recommendation is 0.5, 0.25 and 0.25, respectively. 4.
This section presents the results from the analysis of the ratings made by the subjects of the experiment. The success of a recommendation model to discover new and relevant tracks for a subject can be evaluated without further user input.
A measurement of the success of a recommendation model to discover new and relevant tracks for a subject is the ratio of already discovered tracks and all rated tracks. Figure 3 represents how these sets relate for a particular user.
New Rated
Relevant
Histograms are used to compare distributions of ratings. The x axis represents number of stars/rating and the corresponding percentage of ratings on the y axis. Heat maps are used to show how the ratings of a subject relates to ratings on the same tracks from friends and non-friends. The ratings for social and non-social recommendations of known and unknown tracks are compared, showing that social recommendations give better ratings than non-social ones.
Figure 4 shows the ratings for the tracks known to the subject and for those speci ed as unknown, respectively represented as black and white bars. This shows a clear distinction between known and unknown tracks, most unknown tracks are disliked whereas most known tracks are liked. In this gure we looked indi erently at all the recommenders, in the next gure we di erentiate between collaborative and random recommendations.
Figure 5 represents the proportions of ratings for collaborative and random recommendations. Collaborative recommendation outperforms random recommendation as 45% of unknown known
Col aRbaonradtoivme 2
3
Ratings 0 1 4 5 0 1 2
3 Ratings 4 5 its recommendation get 3 star or above ratings. The percentage drops to around 17% in the case of random recommendations. In this visualization, an increased number of ratings above the threshold of 3 stars is evident for collaborative recommendations. This is due to the way collaborative recommendations are made: only tracks which get a rating superior to 2 stars from other users are used as collaborative recommendations. This shows that there is some consistency in users' ratings. We now look only at the collaborative ratings dividing them between social and non-social recommendations.
Figure 6 represents the ratings from two types of collaborative recommendations. The social and non-social ratings have di erent distributions indicating that subjects react di erently to recommendations coming from their friends than from people they do not know. The subjects are not aware of the source of the recommendation. More than 47% of social recommendations get 3 or more stars as compared to the non-social ones, (33%). In this visualization, we see that social recommendations tend to lead to better ratings, but we mix known and unknown tracks.
Figure 7 represents the ratings of collaborative recommendations on known and unknown tracks. 94% of the ratings state that the track is unknown to the subject, which is why the left histogram on Figure 7 is similar to the histogram of Figure 6. In general, the social recommendations lead to more unknown good recommendations (46% at least 3 stars against 34%) than the non-social ones and less bad recommendations (40% at most 2 stars against 60%). The histogram on the right shows ratings where the subject speci ed she knew the track. Although, the overall rating distributions are very di erent, the same trend of higher ratings for social recommendations is evident.
Figure 8 shows two heat maps. Each square represents the proportion of ratings made by all users of the application on the same tracks for each rating. The left hand heat map represents the relation between ratings of people who are not friends and the right one the relation between ratings of people who are friends.The contrasted regions on both heat map shows that people agree on ratings on the same tracks. Interestingly, the social heat map is more contrasted, showing that friends agree more with their ratings than people who are not friends. 5.
Our results show that people tend to prefer the music that their friends prefer. One of the limitations of this work is that it is based on the assumption that the relationships in SNS or WMC are with real life friends and family which might not be true for some subjects. Therefore, they might not be very good source for music recommendation or discovery in that case, as they might not share user's music taste. Results suggest otherwise.
Studies on music behaviour also suggest that sometimes
youngsters try to distance themselves from the previous
generation [
The analysis of the results of the experiment on social di usion. lead us to the following conclusions: 50 40 30 20 10 0 50 40 30 20 10 0
These conclusions support the view that social di usion is a good mechanism for music recommendation and discovery. It is anticipated that it shall form the foundation for the framework of better music recommender systems combined with social media.
An improved recommendation system could make use of the genres (or tags) of the tracks previously rated by the subject. We have been working on a di erent input for music videos. We are now fetching the Youtube videos posted on Facebook by registered people to the Starnet application and their friends. This changes the settings of the application and requires us to focus on interaction mechanisms for people to explore a social network. The new application shall allow users to add and remove social contacts to their network, to tune their personalised social radio. Another application of interest is a music recommender system that recommends music based on the type of event and music preferences of the people attending it. The initial prototype of the system is available as a Facebook Application7. People post events (such as party, wedding, etc.) on SNS. Analysis of music taste of the attendees of the events can enable the event organiser to make better decisions on what type of music shall be enjoyable for most of the attendees. However, event categorisation and determination of suitability of music within that particular category can be a challenge. 7http://apps.facebook.com/music_valley/