Automatic Generation of Social Tags for Music Recommendation (Poster ID M70)
Douglas Eck, Paul Lamere, Thierry Bertin-Mahieux & Stephen Green (1) A Boosting-based method for predicting labels such as social tags from audio features (II) Generalizes to poorly-tagged or untagged music (III) Works on relatively large dataset: millions of data-mined tags / ~100,000 mp3s Audio and graphic: shortest path through tag space from Beethoven to The Prodigy.
Path from Ludwig van Beethoven to The Prodigy
trip-hop synthpop soul singer-songwriter rock reggae rap punk progressiverock progressivemetal pop piano newwave metal jazz instrumental indie hip-hop grunge gothic funk folk femalevocalists experimental emo deathmetal dance country classicrock classical chillout britpop british blues ambient acoustic 80s 0.0 2.4

1.6

0.8

Autotag

-0.8

-1.6

-2.4

-3.2

Tags

-4.0

d Lu

wig

van

o eth Be

ven

an Em

ue

lA

x

Be

r xA au

ts

Tri

o ou B ar

rah

em om a

An

e sN

wm

an ae l

Th

ch Mi

c Gia

ch

ino eD

Th

d ea

Te

xan ler ie

a Str

tiq

ue

Lo

l sci a

Ga

De

p om dC

ose

rs

Ro

e am yC

e mb

r

Or

bit

al t Ea

Sta

tic Th

ro eP

dig

y

A Time (total 45 sec.)rtist