from web to map: exploring the world of music

18
Distribute d Computing Group From Web to Map: Exploring the World of Music Olga Goussevskaia Michael Kuhn Michael Lorenzi Roger Wattenhofer Web Intelligence 2008 Sydney, Australia

Upload: lindsey

Post on 23-Feb-2016

41 views

Category:

Documents


0 download

DESCRIPTION

From Web to Map: Exploring the World of Music. Olga Goussevskaia Michael Kuhn Michael Lorenzi Roger Wattenhofer Web Intelligence 2008 Sydney, Australia. Music in the old da ys. Storage media Vinyl records Compact cassetts Compact discs - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: From Web to Map: Exploring the World of Music

DistributedComputing

Group

From Web to Map: Exploring the World of Music

Olga GoussevskaiaMichael KuhnMichael LorenziRoger Wattenhofer

Web Intelligence 2008Sydney, Australia

Page 2: From Web to Map: Exploring the World of Music

2Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 2

• Storage media– Vinyl records– Compact cassetts– Compact discs

• An Album is stored on a single physical storage medium– Sequence of songs given by album– Album is typically listened to as a whole

Music in the old days

organization by album

Page 3: From Web to Map: Exploring the World of Music

3Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 3

Music today

• Huge offer, easily available – filesharing, iTunes, Amazon, etc.

• Large collections– The entire collection is stored on

a single electronic storage medium

– Organization by albums (and other lists) is no longer appropriate

organize by similarity!

Page 4: From Web to Map: Exploring the World of Music

4Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 4

Overview

• Define music similarity

• From Perception to Web– Build a graph of songs

• From Web to Map– Embed the graph into

Euclidean space

• Application prototype: www.musicexplorer.org

Page 5: From Web to Map: Exploring the World of Music

5Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 5

Music Similarity

• Audio content analysis• Metadata analysis• Collaborative filtering

– “people who listen to this song also listen to that song”

Similar or different???

Page 6: From Web to Map: Exploring the World of Music

6Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 6

From Perception to Web• Data from last.fm (20M users)

– Top-50 lists (290K lists, 1.5M distinct songs)– Co-occurrence analysis (normalization

cosine(si,sj)=nij/(ninj)1/2)– 1012 (O(TB)!) pair-wise similarity values

• Building a graph G– Edge weight w(si,sj) = 1/cosine(si,sj)– Sparsening: co-occ ≥ 2, w(si,sj) ≥ threshold– sim(si, sj) = length(shortestPathG(si, sj)) – Still n = 430K, m = 6.3M, and ever growing

• How to operate on G? (assuming G is sparse: m=O(n logn))

– Shortest path computation cost: O(m+logn)=O(n logn)– Memory needed to retrieve one value sim(si, sj):

O(m)=O(n logn)

Order of seconds on a state-of-the-art PC!

Need to store the whole G, even if I only have 50 songs in my collection!

Page 7: From Web to Map: Exploring the World of Music

7Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 7

From Web to Map• Embedding: map vertices of G into points in Euclidean

space, s.t. dG/dE (stretch) is “minimized”.

• Computation cost of sim(i,j): O(1) time, O(1) memory per item

• Embedding algorithms:– Multi Dimensional Scaling (MDS): O(dn2)– Spring embedding (Fruchterman-Reingold): O(n2 + m)– MIS-filtering: O(n log2 Δ)– High-dimensional embedding: O(nl2 + lm)– Landmark MDS (LMDS): O(nld + l3)

– Adaptive computation/quality tradeoff– Suitable for dynamic settings

Page 8: From Web to Map: Exploring the World of Music

8Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 8

Iterative Embedding• Assumption: some links erroneously shortcut certain paths

E [# random edges] = X

• Repeat (X / f) times– embed G (using e.g. LMDS)– Remove (from G) fraction f of edges with highest stretch dE/dG

• Example: Kleinberg graph (20x20 grid, f = 0.003)

Spring embedding output

After 6 rounds After 12 rounds After 30 rounds

Page 9: From Web to Map: Exploring the World of Music

9Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 9

Evaluation

• Music Taxonomy (www.allmusic.com)– Control set: 7K songs with genre information

Genre distance dS= LCA (least

common ancestor)

How well does the resulting map represent music similarity?

Page 10: From Web to Map: Exploring the World of Music

10Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 10

Evaluation: Quality Measures

• Distance comparison QL: average similarity increase as a function of genre distance ds

• Embedding smoothness QR: average # of genre re-occurrences on a random line

Avg. similarity of pairs (si,sj) w/ ds(i,j)=h

Songs that belong to distant genres should be far away in the embedding.

Genre transitions in the embedding should be “smooth”.

Page 11: From Web to Map: Exploring the World of Music

11Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 11

Evaluation: Iterative Embedding

After 30 rounds, f=0.5%LMDS output

(430K nodes, 10 dimensions)

Page 12: From Web to Map: Exploring the World of Music

12Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 12

Evaluation

Closest neighbors

in 10D

Page 13: From Web to Map: Exploring the World of Music

13Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 13

Applications: Music Explorer

• www.musicexplorer.org– Web service to query

coordinates (current DB with 430K titles)

– Visualization in 2D– Zoom level according to

song popularity– Playlist generation based on

trajectories

Page 14: From Web to Map: Exploring the World of Music

14Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 14

Blues

Rock

Text

START

END

Blues-Rock

Electric Blues

Playlist generation• Interpolation between start and end-point

– Smooth transition from one style to the other– In reality: 10 dimensions

Page 15: From Web to Map: Exploring the World of Music

15Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 15

Music in Euclidean Space

• Performance– Similarity computation comes almost for free: O(1) time– Memory footprint is extremly low: O(1) per song

– All information can be saved in the file, no server connection required.

• Applications– Trajectories (playlists, ...)– Volumes (region of interest, ...)– Notion of direction

coordinates are well suited for mobile applications

coordinates are well suited for similarity based organization

Page 16: From Web to Map: Exploring the World of Music

16Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 16

Towards a new world of music?• Euclidean representation

– Efficient similarity computation (time and memory)– No server needed: distributed applications

– Building blocks for new functionalities:

• New scenarios:– Mobile file sharing– P2P overlay based on the map– Innovations at home

– “Play anything hip-hip… not this and not closely related songs… go towards Detroit house, be there in an hour”

– Automatic DJ (collect feedback from mobiles, generate playlists based on guests regions of interest)

Trajectories(Playlists)

Volumes(Interest Regions)

Notion of Direction(Browsing)

Page 17: From Web to Map: Exploring the World of Music

17Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 17

Conclusions

• Necessary?

Page 18: From Web to Map: Exploring the World of Music

18Olga Goussevskaia, ETH Zurich @ Web Intelligence 2008 18

Thanks for your Attention

• Questions?