Slide 1

Michael Kuhn

Distributed Computing Group (DISCO)ETH ZurichThe MusicExplorer Project:Mapping the World of Music

Today, I woud like to listen to something cheerful.Something like Lenny Kravitz would be great.Who can help me to discover my collection?

In my shelf AC/DC isnext to the ZZ Top...

play random songs that match my moodWhats the Talk about?

basic idea: map of music

constructing the map(MDS, PLSA)

using the map(Youtube, Android)Map of Music: What is it good for?Similar songs are close to each other

Quickly find nearest neighbors

Span (and play) volumes

Create smooth playlists by interpolation

Visualize a collection

Low memory footprintWell suited for mobile domainAdvantages of a MapHey JudeImagineMy PrerogativeI want it that wayPraise youGalvanizerockpopelectronicconvenient basis to build music softwareHow to Construct a Map of Music?

Similar or different???

Music SimilarityAudio Analysis

Usage Data

From Usage Data to Similarity

Collaborative Filtering

Folksonomies (Tags)Collaborative Filtering and MDSMethod 1:Basic Idead = ?item-to-item collaborative filtering1graph for all-pairsdistances2MDS to embed graph (i.e. distances)into Euclidean space3Item-to-Item Collaborative Filtering

People Who Listen To This Song also Listen to...[Linden et al., 2003]

Users who listen to A also listen to BTop-50 listened songs per userNormalization (cosine similarity)

Pairwise Similarity

#common users (co-occurrences)Occurrences of song AOccurrences of song BGraph EmbeddingEBDCA2334545BAEDCWell-known for dimensionality reductionfirst described by Young and Householder, 1938

Principal Component Analysis (PCA): Project on hyperplane that maximizes variance.Computed by solving an eigenvalue problem.

Basic idea of MDS:Assume that the exact positions y1,...,yN in a high-dimensional space are given.It can be shown that knowing only the distances d(yi, yj) between points we can calculate the same result as applying PCA to y1,...,yN.

Problem: Complexity O(n2 log n)Classical Multidimensional Scaling (MDS)Select k landmarks and embed them using MDS

For the remaining points: Place according to distances from landmarks

Complexity: O(k n log n)Landmark MDS[de Silva and Tenenbaum, 1999]Assumption: some links erroneously shortcut certain paths

Idea: Use embedding as estimator for distanceShortcut edges get stretchedRemove edges with worst stretch and re-embed

Example: Kleinberg graph (20x20 grid with random edges)Iterative Embedding

Original embedding(spring embedder)After 6 roundsAfter 12 roundsAfter 30 rounds19Kleinberg graph: grid edges + one random edge per node (follows 1/d^2 distribution, d=manhattan distance)Evaluation: Dimensionality

Pink Floyd - TimePink Floyd - On the RunPink Floyd - Any Colour you LikePink Floyd - The Great Gig in the SkyPink Floyd - EclipsePink Floyd - Us and ThemPink Floyd - Brain DamagePink Floyd - Speak to MePink Floyd - MoneyPink Floyd - BreathePink Floyd - One of These DaysMiles Davis - So WhatHorace Silver - Song For My FatherBill Evans - All of YouMiles Davis - Freddie FreeloaderNat King Cole - The More I See YouMiles Davis - So NearMiles Davis - Flamenco SketchesCharles Mingus - Eat That ChickenJimmy Smith - On the Sunny SideJulie London - DaddyBill Evans My Mans Gone Now10 Dimensions give a reasonable qualityExample Neighborhoods in 10D Space20Dimensionality: - quality(distorsion)/space(performance) tradeoff - max QR at 10D, in contrast to other work, which focuses on at most 3D (visualization) - The axis are not assigned any semantical meaning (but maybe one could find that out? Open problem)

Neighborhood lists: search a little on whether the jazz list is really good.Social Tags and PLSAMethod 2:

Similarity from Social Tagging

Probabilistic Latent Semantic Analysis (PLSA)w1w2wMZ1 =?Z2 =?ZK =?P(w|z)P(z|d)d1d2dNdocumentslatent semantic classeswordssongslatent music style classestags[Hofmann, 1999]PLSA: Interpretation as Spacew1w2wMZ1 =?Z2 =?ZK =?P(w|z)P(z|d)d1d2dNsongslatent music style classestagscan be seen as a vector that defines a point in space [Hofmann, 1999]K small: Dimensionality reductionPLSA Space

latent class 2latent class 3latent class 1Probabilities sum to 1:K-1 dimensional hyperplaneSimilar documents (songs) are close to each othermusic space: 32 dimensionsAdvantages of LMDS:Same accurracy at lower dimensionality (10 vs. 32)

Advantages of PLSA:Natural meaning of tagsAssignment of tags to songs (probabilistic)LMDS vs. PLSA SpaceCurrent sizes (approx.):LMDS: 600K tracksPLSA: 1.1M tracksUsing the MapVisualization?

high-dimensional!Identify relevant tags

Find centroids of these tags in 10D

Apply Principal Component Analysis (PCA) to these centroidsFrom 10D to 2D

What people have chosen during the researchers night in ZurichYouJuke The YouTube Jukebox

YouTube as media sourceMusic map to create smart playlist33

34

Reaching YouJukewww.youjuke.org

apps.facebook.com/youjuke

seek

play random songs that match my mood

In my shelf John Lennon isnext to the Beatles...seek:The map of music on Android

An Intelligent iPod-Shuffleskip = listen = RealizationAfter only few skips, we know pretty well which songs match the users moodWork in Progress: Who is Dancing?

AC/DCBeatlesProdigy

In my shelf AC/DC isnext to the ZZ Top...Browsing Covers

Video

www.musicexplorer.org/museekInternet connection only required at first startup!Thanks to:Lukas Bossard Mihai CalinOlga GoussevskaiaMichael LorenziRoger WattenhoferSamuel Welten

URLs:www.musicexplorer.org/museekwww.youjuke.orgapps.facebook.com/youjuke

E-Mail:kuhnmi@tik.ee.ethz.ch (Michael Kuhn)Questions?