Music Recommendation Systems: A Progress Report

1
Music Recommendation Systems A Progress Report

• Adam Berenzweig
• April 19, 2002

2
Music Recommendation Is

• Music IR for the masses
• Kids in candy stores
• Querying is hard people cant describe music
• Recommendation can be integrated into players,
  streaming services, music stores, etc.
• Break major label/retail monopoly on choice!!

3
Music Recommendation Is

• Set-based IR
• Find me items similar to this set, in the way
  that the set is similar to itself
• Set collection, or playlist extension.
• Be sensitive to themes or aspects of the users
  collection.
• All about similarity

4
Background I IR/Statistics

• Collaborative Filtering
• Latent Semantic Analysis (Deerwester al., 90)
• SVD to find hidden meaning
• Probabilistic LSA (Hofmann, 99)
• EM to find hidden meaning
• Latent Class Models (Hofmann, 99)

5
Background II Audio IR

• Artist classification
• Whitman Lawrence Berenzweig, Ellis Lawrence
• Genre classification
• Tzanetakis
• Fingerprinting, query-by-example
• What features???
• What is it I like about the music that I like?

6

• Artist Classification Using Vocals
• Anchor Models
• Similarity Metrics

7
Artist Classification Using Vocals

• Are vocal segments more easy to identify than
  instrumental segments?
• Using Voice Segments to Improve Artist
  Classification of Music, Berenzweig, Ellis
  Lawrence, to appear AES 22.

8
Segmented Posteriograms
9
Segmented Posteriograms
10
Experiment at-a-glance
Frame labels
Song labels
Audio input
Cepstra (MFCC, PLP)
Artist Classifier
Confidence Weighting
Vox/Music Classifier
11
Results
12
The Album Effect

• Testing on different album than trained hurts
  performance by 30-40 relative.
• Is it production effects or style?

13
Future Work

• Album Effect production or style?
• Better segmentation
• Further analysis of posteriograms
• song structure change detection, clustering
• another level of classification? leads to...

14

• Artist Classification Using Vocals
• Anchor Models
• Similarity Metrics

15
Anchor Models

• Dual Motivation
• Scalable artist classification
• Induced artist similarity metric
• Technique from speaker identification literature
  (Reynolds, Sturim al.)

16
Anchor Models
Anchor Models
17
Anchor Space

• n-dimensional Euclidean space
• Distance metric is simple
• Dimensions have meaning

18
Anchor Models

• Basically doing dimensionality reduction or
  feature extraction, where
• nonlinear mapping to low-D feature space is
  learned
• mapping is musically relevant
• but no theoretical justification like PCA

19
Anchor Space

• Artists are distributions, not points.
• Model with GMMs
• Each frame of audio (32 milliseconds) is a point.
• Each song is a cloud, too.
• Distance is KL-divergence
• estimate with total likelihood under GMM.

20

• Artist Classification Using Vocals
• Anchor Models
• Similarity Metrics

21
Searching for Ground Truth

• Does a single correct similarity metric exist?
• Subjective, relative, mood-dependent.
• Aspects of similarity - Tversky 77
• (Psychological) similarity is not a metric.
• A dynamic interplay between classification
  similarity

22
Similarity is not a metric?
An ellipse is like a circle. A circle is like an
ellipse.
No Triangle Inequality
Asymmetry
23
Salient Aspects

• Distance in big Euclidean space may not have any
  meaning!
• Goal find big Euclidean space, then analyze
  salient dimensions of collections.
• Directly answers the question what is it I like
  about the music that I like?

24
Searching for Ground Truth

• Sources of Opinion
• Ask directly?
• Preference Data Spidering opennap lists.
• Expert Opinion Allmusic Guide Similar Artist
  sections.
• Semantic Similarity Whitman Lawrence

25
Semantic Similarity

• Community Metadata. (Whitman and Lawrence)
• Web spider collects terms.
• Treats artists like documents

26
Expert Opinion
27
Completing the Erdos Numbers
Incomplete Graph
Complete Erdos Distance
28
Human Evaluation

• Want many judgements, but full matrix not likely
• Problem of relativity, drift
• Ask for relative judgements
• Game and Survey mode
• Problem of unknown artists
• Use total history

29
Musicseer
30
Evaluation Ranking
31
Thanks!