Music Recommender Systems

longtermagonizingInternet and Web Development

Dec 13, 2013 (5 years and 26 days ago)


Modeling Users' Intentions for the Enhancement of
Music Recommender Systems

Asma Rafiq


Centre for Digital Music




Research Questions

Research Plan




Problem: Online music stores offer millions of
songs to choose from, users need assistance.

Solution: Music
Recommender Systems

Using social networks, ontologies and the
semantic web, this research work seeks to
assist people in finding the music they might
be interested in.



The Web is leaving the era of

entering one of

What's the difference?

Search is what you do when you're looking for

Discovery is when something wonderful that you
didn't know existed, or didn't know how to ask for,
finds you.”


Recommender Systems

Recommender systems are software applications that aim to
support users in their decision
making while interacting with
large information spaces. They recommend items of interest
to users based on preferences they have expressed, either
explicitly or implicitly.

Recommender systems help overcome the information
overload problem by exposing users to the most interesting
items, and by offering novelty, surprise, and relevance.

Recommender technology is hence the central piece of the
information seeking puzzle. Major music services such as are using recommendation technology in ubiquitous


Recommendation Strategies

Collaborative Filtering
is a widely used approach to solve the
recommendation problem. The stored interaction (explicit or
implicit) between the users of the system and the item set
helps generate informed guesses for recommendations.

based Filtering
collects the information regarding the
items and based on user preferences filters the results that
the user is most likely to prefer. It simply depends on item
description rather than the user ratings.

based Filtering
uses the contextual information to
describe the items.

Demographic Filtering
results based on stereotypes of
users that like certain item.

Hybrid Filtering
is simply the combination of two or more
recommendation strategies.


Music Recommender System

Music is inherently different than other types of media. The
space of recommended items is extremely large as compared
to other domains .

People interact with music differently than they do with other
types of media e.g. repetitive listening.

Listeners vary their music preference based upon context and

Listeners enjoy listening to sequences of songs often getting
as much enjoyment from the song transitions as from the
songs themselves.

It is important to consider the special nature of music when
building recommenders for music.

The uniqueness
of music as recommendation domain present
challenges not seen in other
recommender domains.


Research Questions

What are the resources that can be used to model the intentions of
the users for recommendation strategies?

Profiles are very important resource to recommend music to the users, people
prefer the genres that their friends and family listen to.

Profiles of people outside of one’s own network might also play role in music
discovery as some users indicated in Music Valley application feedback.

Music listed in Social Network websites, might not reveal the taste of the user,
as many other factors might be influential on listed preferences e.g. band
promotion requests by acquaintances, which might not actually be listened by
the users.

The users age is important in evaluating user music preferences, also the
system restricts the profile to above 18, the users mostly lied in the range of
45 years of age.

The education level for most of the users was post graduate. Indicating that
the system was mostly used by highly educated people.


Profiles of users reveal important information such as demographic
information, interests, event information, etc. that can be used to
design recommender models

What are the ethical implications involved in acquiring these

The ethical implications were carefully analysed for the experiment

Subjects under the age of 18 should not be allowed access to the
system due to strict laws in various countries regarding underage
teens and children.

Some of the subjects when requested for granting the permissions for
using the application denied the permissions which was required for
the data extraction from their profiles.

It seems some social network users have high privacy concerns.

Some users suggested to explain the application functionality first
before asking for the permissions. However, the links to privacy
statement and terms of services were

ignored by the users.


What is the state of the current music recommender systems
and how to improve it using the resources identified as useful
in user
intention extraction?

Music recommender systems have limited access to user information
spread all over the web, linked information can help model user taste
more accurately.

Recently, some music recommender systems incorporated the
importance of social aspect of music introducing features like playlist
sharing, commenting on friend’s playlists, etc.

Social network profiles are identified as useful resource for music
recommendation. It could be used to creating a trust based music
recommender system with the identified music taste leaders.

Music recommender systems might improve results by adapting to the
users current situation.

The existing recommendation techniques such as content
collaborative fil
tering or hybrid techniques focus on users explicit
contact behaviours
but ignore the implicit relationship among users in
the network.


Has semantic web reached the level of maturity to develop
the anticipated sophisticated music recommendation

To answer this question a detailed analysis of current semantic web is

I shall conduct this in the next month which is also required to design
the next phase

Knowledge base could be established that contains information as

How can we evaluate the success of user modelling in
recommender system?

The evaluation of the system can be done by comparing the social
music recommender system and semantic music recommender system

The user’s ratings and feedback will be used for this purpose




Phase 1: Social Music Recommender Systems

The experiment will be finalised by the end of this month and results
and feedback on the experiment shall be helpful in drawing various
conclusions and proposing future work.

Research paper accepted in August 2011

Addressed research questions related to music discovery and
recommendation using online social networks

In this paper, I have drawn conclusions on the present literature
regarding social networks, web music communities, how such systems
can be useful and proposed future work

ready version due on 19 September 2011

I will be presenting this paper in WOMRAD in conjunction with ACM
RecSys 2011 on 23 October 2011 in Chicago, IL


Facebook Application

A social music recommender system is a system that
extends on social interaction and events

Identify user’s event posted on Facebook

Extraction of this information through the Facebook Graph

Recommend music based on the preferences of
participants of the event

Suggest the most popular songs using Youtube



User Feedback

The user was surprised to see her favorite songs being recommended to
her, I ellaborated the experiment and she told me that the bands/artists
she listed in her preferences (a total of 8) were


the bands she actually
listened to, rather they were some friends whom she was asked to
promote by ‘liking’ their fan page. This revealed that the
recommendations generated from friends' preferences were leading to
more useful results for this user.

Saves time for a user to think of what she wants to listen next (as the
songs are recommended as a playlist).

The system has been successful in re
discoveries of songs by
musician/bands that a user admired while she was a teen but had
forgotten about.


The usual Youtube recommendations did not interest a user as she said: “I
never liked the recommendations

it is always what Youtube want to sell
me i.e. the songs by the song owners /publishers

who always have
advertisements at the start of the clip. So, I never click on these
recommendations but this is much better as a player because its playing
the songs I know, and want to listen to and not the ones that will show ads

Sarcastic appreciation: “This system is creepy! I have been recommended
a song from a band that I used to login to Facebook account!! I smell
conspiracy theory here...”

“The application asks for many permissions; an explanation before a user
is asked to grant the permissions would be helpful to know what is going

“The next stage for music recommendation people could also pick the
quality they like to listen to i.e., only choose clips that have HD sound or
don't pick the ones that are kids playing in their bedroom!”

The songs should be mixed within the playlist from different artists, it
makes it boring to use the system for a randomly picked artist playlist.


Problems with Social Platforms

Change of backend support without prior notice

FBML and RestAPI are deprecated

Change of interface without prior notice

The application is displayed in a canvas on the Facebook platform. The
canvas is actually an IFrame in HTML. An IFrame is a visual part of a
Web page which contains another Web page. When the Facebook
page of the application is rendered on the client, the IFrame interacts
with our server to authenticate the user and the application.

Now, the allocated space is also reduced which used to be of 760
pixels wide.

Fetching friends data using the Graph API is quite slow (takes up to 20
seconds for one user)

Even the display of Facebook users in the application is slow (getting
their name and picture).

All in all working with the Facebook API is problematic.


Phase 2: Semantic Music
Recommender System

One good application is to illustrate the effect the semantic
Web can have on music recommendations, once data has
become semantically structured.

We are going to look at music ratings as a semantically rich
information source that when related to music files has many
advantages and creates a whole new Web of potential

user similarity is a fundamental component of
Collaborative Filtering (CF) recommender systems. In user
user similarity the ratings assigned by two users to a set of
items are pairwise compared and averaged (correlation).


In this experiment we intend to make user
user similarity
adaptive, i.e., we dynamically change the computation
depending on the profiles of the compared users and the
target item whose rating prediction is sought.

We propose to base the similarity between two users on the
subset of co
rated items which best describes the taste of the
users with respect to the target item.

These are the items which have the highest correlation with
the target item.

We will evaluate the proposed method to show that the
proposed locally adaptive neighbor selection, via item
selection, can significantly improve the recommendation
accuracy compared to standard CF.


Semantic Music Recommender System is a personalised music
recommender system which tries to limit the problems of
collaborative recommender systems by ontologically using
semantic information from the categorical characteristics of
an item such as Genre.

The similarities between user pairs will be calculated by a
weighted mean method that calculates three similarity

The similarity of user evaluation histories (using the Pearson
correlation coefficient on usage information of the system in terms of
a user
item evaluation data);

The similarity of these user's demographic data

The users similarity in interest or preference based on the semantic
similarities of the items retrieved and/or evaluated.


The use of ontologies in these types of systems limits specific
problems, including the following:

To guarantee the inter
operability of system resources and the
homogeneity of the representation of information.

To allow for the dynamic contextualisation of user preferences in
specific domains.

To facilitate performance in social networks and collaborative filtering.

To improve communication processes between agents and between
agents and users.

To limit the "cold start" problem by completing the incomplete
information through inferences.

The ability to semantically extend descriptions of user contextual

To improve the representation and description of different system

Improve the description of system's logic by admitting the inclusion of
a set of rules.

Provide the necessary means to generate descriptions enriched by
web services and facilitate their discovery by software agents.


Dataset Candidates

Dataset 1: Yahoo! Music User Ratings of Musical Artists, version 1.0

This dataset represents a snapshot of the Yahoo! Music community's
preferences for various musical artists.

The dataset contains over ten million ratings of musical artists given by Yahoo!
Music users over the course of a one month period sometime prior to March

Users are represented as meaningless anonymous numbers so that no
identifying information is revealed.

The dataset can be used by researchers to validate recommender systems or
collaborative filtering algorithms.

Dataset 2: Million Song Dataset


is a freely
available collection of audio features and metadata for a million
contemporary popular music tracks.

The dataset does not include any audio, only the derived features.


Phase 3: User Modeling in Music
Recommender System

In the long
run, I intend to design a framework that
incorporates the best and possible integration of these two
intention modelling perspectives (social and semantic) in
order to model the intention of the user effectively.


Gantt Chart of Future Plans



People need assistance to search the needle in the haystack in this era of
information overload.

The best way to achieve this is to make our computing machines ‘aware’ of the user

The semantic and intentions gaps need to be addressed in order to
achieve satisfactory results.

It is the need of time to review the enhancements in the recommendation
systems with resources available that could be extracted from current
massive media to model the user intentions in order to improve the music
experience for a wide range of people; leading to the development of a
improved music recommendation system


Questions and Comments