This paper presents Myusic, a platform that leverages social media to produce content-based music recommendations. The design of the platform is based on the insight that user preferences in music can be extracted by mining Facebook pro les, thus providing a novel and effective way to sift in large music databases and overcome the cold-start problem as well. The content-based recommendation model implemented in Myusic is eVSM [4], an enhanced version of the vector space model based on distributional models, Random Indexing and Quantum Negation. The e ectiveness of the platform is evaluated through a preliminary user study performed on a sample of 50 persons. The results showed that 74% of users actually prefer recommendations computed by social mediabased pro les with respect to those computed by a simple heuristic based on the popularity of artists, and con rmed the usefulness of performing user studies because of the di erent outcomes they can provide with respect to o ine experiments.
One of the main issues of the so-called personalization pipeline is preference
acquisition and elicitation. That step has always been considered the bottleneck
in recommendation process since classical approaches for gathering user
preferences are usually time consuming or intrusive. The widespread di usion of social
networks in the age of Web 2.0 o ers a new interesting chance to overcome that
problem, since users spend 22% of their time on social networks3 and 30
billion pieces of content are shared on Facebook every month [
This paper presents Myusic, a tool that provides users with music
recommendations. The goal of the system is to catch user preferences in music and
lter the huge amount of data stored in platforms such as iTunes or Amazon
in order to produce personalized suggestions about artists users could like. The
ltering model behind Myusic is eVSM, an enhanced extension of VSM based on
distributional models, Random Indexing and Quantum Negation. As introduced
in [
Generally speaking, this work concerns to the area of music recommendation.
The commonly used technique for providing recommendations is collaborative
ltering, implemented in very well known services, such as MyStrands4, Last.fm5
or iTunes Genius. An early attempt to recommend music using collaborative
ltering was done by Shardanand [
The paper is organized as follows: the architecture of the systems is sketched in Section 2; Section 3 focuses on the results of a preliminary experimental evaluation and nally Section 4 contains conclusions and directions for future research. 2
The general architecture of Myusic is sketched in Figure 1. We can identify four main components:
Crawler. The Crawler module queries Last.fm through its public APIs to build a corpus of available artists. For each artist, the name, a picture, the title of the most popular tracks, their playcount and a set of tags that describe that artist are crawled. All the crawled data are locally stored.
Extractor. The Extractor module connects to Facebook, extracts artists the user likes (Favourite Music section in the Facebook pro le, see Figure 2), and maps them to the data gathered from Last.fm in order to build a preliminary set of artists the user likes. This information is locally modeled in her own pro le to let her receive recommendations even in her rst interaction with Myusic, thus avoiding the cold-start. Implicit information coming from the links posted by the user and the events she attended are extracted, as well.
Pro ler. The process of building user pro les is performed in two steps.
First, a weight is assigned to each artist returned by the Extractor. The
weight of a speci c artist is de ned according to a simple heuristic: if a user
posted a song, that information can be considered as a light evidence of her
preference for that artist, while the fact that she explicitly clicked on "Like"
on her Facebook page can be considered as a strong evidence. For example,
on a 5-point Likert scale, a score equal to 3 is assigned to the artists whose
name appear among the links posted by the user, while a score equal to 4 is
assigned to those occurring in her favorite Facebook pages. If an artist occurs
in both lists (that is to say, the user likes it and posted a song, as well), 5 out
of 5 is assigned as score. Next, a pro ling model has to be chosen. The eVSM
framework provides four di erent pro ling models [
Recommender. Given a semantic vector space representation based on distributional models for both artists and user pro les, through similarity measures it is possibile to produce as output a ranked list of suggested artists. The cosine similarity for all the possible couples (pu; a) is computed, where pu is the vector space representation of user u, while a is the vector describing the artist a. Figure 3 shows an example of recommendation list. The platform allows the user to express feedbacks on recommendations. Positive and negative feedbacks are used to respectively update positive and negative pro le vectors and to trigger the recommendation process again. 3
The goal of the experimental evaluation is to validate the design of the platform by carrying out a user study whose goal is to analyze the impact and the e ectiveness of the di erent con gurations of eVSM implemented in Myusic. Speci cally, a user study involving 50 users under 30, heterogeneously distributed by sex, education and musical knowledge (according to the availability sampling strategy) has been performed. They interacted for two months with the online version of Myusic. A crawl of Last.fm was performed at the end of November, 2011 and data about 228,878 artists were extracted. Each user explicitly granted the access to her Facebook pro le to extract data about favourite artists. At the end of the Extraction step, a set of 980 di erent artists the 50 users like were extracted from Facebook pages. Generally speaking, 1,720 feedbacks were collected: 1,495 of them came from Facebook pro les, while 225 were explicitly provided by the users (for example, expressing a feedback on their recommendations). The collected feedbacks were highly unbalanced since only 116 (6.71%) on 1,720 were negative. Last.fm APIs were exploited to extract the most popular tags associated to each artist. The less expressive and meaningful ones (such as seenlive, cool, and so on) were considered as noisy and ltered out. The design of the user study was oriented to answer to the following questions: { Experiment 1: Does the cold-start problem can be mitigated by modeling user pro les which integrate information coming from social media? { Experiment 2: Do the users actually perceive the utility of adopting weighting schemes and negation when user pro les are represented? { Experiment 3: How does the platform perform in terms of novelty, serendipity and diversity of the proposed recommendations?
In the rst experiment, users were asked to login and to extract their data
from their own Facebook page. Next, a user pro le was built according to a
proling model randomly chosen among the 4 described above and a preliminary
set of recommendations was proposed to the target user. In order to evaluate
the e ectiveness of the Extractor we compared the recommendation list
generated through eVSM to a baseline represented by a list produced by simply
ranking the most popular artists. Next, we asked users to tell which list they
preferred. Obviously, they were not aware about which list was the baseline and
which one was built through eVSM. A plot that summarizes users' answers is
provided in Figure 4-a. It is straightforward to note that users actually prefer
social media-based recommendations, since 74% of them preferred that
strategy with respect to a simple heuristic based on popularity of the artists stored
in database. However, even if the results gained by this pro ling technique were
outstanding, it is necessary to understand why 26% of the users simply preferred
the most popular artists. Probably, there is a correlation between users'
knowledge in music and the list they choose. It is likely that users with very generic
tastes prefer a list of popular singers. Similarly, it is likely that users with a
poor knowledge in music might prefer a list of well-known singers with respect
to a list where most of the artists, even if related to their tastes, were unknown.
A larger evaluation with users, split according to their musical knowledge, may
be helpful to understand the dynamics behind users' choices. Similarly, it would
be good to investigate the impact of the amount of the information extracted
from Facebook pro les with the accuracy of the recommendations. The second
experiment was performed in two steps. In the rst step users were asked to login
and to extract their data from their own Facebook page, as in Experiment 1.
Next, two pro les were built by following the RI and the W-RI pro ling models,
respectively. Finally, recommendations were generated from both pro les, and
users were asked to choose the con guration they preferred. As in Experiment 1,
they were not aware about which recommendations were generated by
exploiting their weighted pro le and which ones were produced through its unweighted
counterpart. Results of this experiments are shown in Figure 4-b. Di erently
from the results obtained from an in-vitro experiment performed in a scenario
of movie recommendation [
In this paper we proposed Myusic, a music recommendation platform. It
implements a content-based recommender system based on eVSM, an enhanced
version of classical VSM. The most distinguishing aspect of Myusic is the
exploitation of Facebook pro les for acquiring user preferences. An experimental
evaluation carried out by involving real users demonstrated that leveraging
social media is an e ective way for overcoming the cold-start problem of CBRS. On
the other hand, the exploitation of relevance feedback and user ratings generally
did not improve the predictive accuracy of Myusic. Users showed to trust the
system, and Myusic also achieved good results in terms of accuracy and diversity
of recommendations. Those results encouraged keeping on this research. In the
future we will investigate the adoption of recommendation strategies tailored on
the music background of each user, even by learning accurate interaction models
in order to classify users [