Martin Przyjaciel-Zablocki • Thomas Hornung • Alexander Schätzle

Martin Przyjaciel-Zablocki • Thomas Hornung • Alexander Schätzle

MUSE: A Music Recommendation Management System
Martin Przyjaciel-Zablocki Thomas Hornung Alexander Schätzle Sven Gauß Io Taxidou
Department of Computer Science, University of Freiburg, Germany
Georg Lausen
{zablocki, hornungt, schaetzle, gausss, taxidou, lausen} @informatik.uni-freiburg.de
-
 Music platforms such as Spotify, Google Play Music or
Pandora make millions of songs accessible
MUSE (Music Sensing in a Social Context)
 Music Recommendation Management System
(MRMS) for developing and evaluating
music recommendation algorithms
 Takes care of all regular tasks required for
in vivo evaluations
 Simple API enables support for arbitrary
music recommenders
 Connectors to social networks & semantic
web sources are added continuously
 Open sourced under Apache License 2.0:
 Social Networks and the Semantic Web describe
users and items in an unprecedented granularity
 New opportunities and challenges for
Music Recommenders
Challenges for development
 Prototypes programmed from scratch, again and yet again!
 Immediate feedback available
 Manifold sources of information to combine
Recommendation
List
Administration
Evaluation
Framework
User
Profile
User
Context
XML
Last.fm API
RDF
User Profile Engine
RDF
Freebase
DBpedia
Social
Connector 1
XML
Last.fm
Spotify API
Web Services
Wrapper/
Connector 1
HTML
Charts
FaceBook
Listen & rate recommended songs
PRIVACY &
SECURITY
Variety of baseline algorithms
for comparison included
Users
HTML
JSON
EVALUATIONS
Real-time monitoring
Results are plotted
and can be exported
Compose recommendation
lists out of diverse algortihms
Semantic Web
Track
Manager
Music
Repository
Music Retrieval Engine
AJAX
Spotify
Connector
EXTENSIBILITY
MUSIC
Social Networks
Web-based User Interface
Data Repositories
User
Manager
REST
Webservice
Conforms with
state-of-the-art
privacy standards
Let s stop reinventing the wheel and put the fun
back in developing great new algorithms!
Coordinator
Recommender
Interface
MuSe Client
PLAYING
SONGS
New recommender
algorithms can be
plugged in comfortably
M U SE
muse.informatik.uni-freiburg.de
MuSe Server
Recommendation
List Builder
Play and rate
recommended
songs from Spotify
An public instance of MUSE with a small user
community is online:
Recommender Manager
Recommendation
Model
MANAGEMENT
RECOMMENDERS
Pluggable Recommender 1
Recommender 1
Administrative access
gives control and insign
USER
dbis.informatik.uni-freiburg.de/MuSe
Challenges for evaluations
 Thorough evaluations are repetitive, yet non-trivial, tasks
 Many aspects are inherently subjective, e.g. serendipity
 can be only investigated in vivo (with real users)
 Legal issues: (1) user data is highly sensitive
(2) content-based approaches require data
Recommender
Users provide profile
information or social identities
Compose recommendations
Developers
Web-based User Interface
 Register & interconnect social identities
 Listen & rate recommended songs
 Manage new recommenders & evaluations
Data Repositories
 3 shared data structures (restricted access!)
 MUSIC RETRIEVAL ENGINE collects new songs
and retrieves meta information
 USER PROFILE ENGINE retrieves information
from linked social profiles
Recommender Manager
 Coordinates interaction of recommenders
with users and the data repositories
 Forwardes user request & receives
generated recommendations
 Composes list of recommendations
Case Study
 48 participants
 1567 rated song
recommendations
 Overall, 73% of all
songs positively
rated
Annual Charts
Taste in music evolves between the age
of 14 and 20 [11]
Country Charts
Inter-country view
on songs
Quality score per recommender type
{love  2, like  1, dislike  -1}
Configure & manage evaluations
Browse & analyze results
City Charts
Explore upcoming
city trends
Social Tags &
Social Neighborhood
Social Networks and
Semantic Web as
source of information
dbis.informatik.uni-freiburg.de/MuSe
A Music Recommendation Management System
 MUSE simplifies development and evaluation of music
recommendation algorithms
 Provides many tools for developing recommenders
 Takes care of typical tasks of in vivo evaluations
 Open sourced under Apache License 2.0
Future Work
 Increase flexibility of evaluations
 Further connectors for social networks
 New algortihms for music recommendation