Transcript Classifier ensembles

Stochastic Text Models for Music Categorization
Carlos Pérez-Sancho, José M. Iñesta, David Rizo
Pattern Recognition and Artificial Intelligence group
Department of Software and Computing Systems
University of Alicante, Spain
Outline
► Introduction
► Music
encoding
 Melody
 Harmony
► Experiments
 Plain classification
 Classifier ensembles
 Hierarchical classification
► Conclusions
SSPR 2008
2
Introduction
► Premise:
music content can be used to model
musical style
► We
use language modeling techniques to
classify symbolic digital scores
► For
that, digital scores need to be encoded into
sequences of symbols
SSPR 2008
3
Music encoding
► Two
different sources of information
Melody
Harmony
SSPR 2008
4
Melody encoding
►
►
►
Polyphonic sequences are reduced to monophonic using skyline
Pitch intervals and duration ratios are computed for each pair of
consecutive notes
Numeric values are encoded into ASCII symbols
Bf AZ aF bZ
BZ
dZ bZ DZ
input string
(2,×½) → Bf
(1,×1) → AZ
(-1,×2) → aF
(-2,×1) → bZ
(2,×1) → BZ
( interval , duration ratio )
(-4,×1) → dZ
(-2,×1) → bZ
(4,×1) → DZ
SSPR 2008
5
Harmony encoding
Key: E flat
►
►
VIm
V
I
input string
Chords are encoded as degrees relative to the tonality
for transposition invariance
Only chord changes are encoded
SSPR 2008
6
Experiments
► Dataset:
music from 3 genres and 9 sub-genres
(around 60 hours of music)
► Classification
techniques
 Naïve Bayes
 Language modeling (n-grams), classifying by lowest
perplexity
SSPR 2008
7
Experimental setup
► First
step: plain classification
 80% of the dataset used
 10-fold cross validation
► Second
step: hierarchical classification using
classifier ensembles
 Weights of the ensembles adjusted using previous

results
Remaining 20% of dataset used for validation
SSPR 2008
8
Classification results
Chords (harmony)
Melody
2-grams 3-grams 4-grams N.B. 2-grams 3-grams 4-grams N.B.
►
Best classification rates in the 3-classes problem were obtained
using harmonic information
►
When classifying sub-genres, melody usually performs better
►
Naïve Bayes performs better most of the times
►
No significant differences for different context sizes in n-grams.
SSPR 2008
9
Confussion matrix
×100 %
►
►
Misclassifications occur more frequently within broad domains
Try to prevent intra-domain errors by using a hierarchical classifier
SSPR 2008
10
Hierarchical classification
SSPR 2008
11
Hierarchical classification
►
Harmony (chord progressions) is used at the first level
►
Melody is used at the second level
►
Instead of using single classifiers, an ensemble of
classifiers is used at each level to increase robustness
SSPR 2008
12
Classifier ensembles
►
►
Decisions are made by weighted majority vote
Two weighting schemes
 Linear best-worst weighting vote
 Quadratic best-words weighting vote
weight
# errors
# errors
SSPR 2008
13
Hierarchical classification results
►
With the remaining 20% of the dataset
Single classifiers
2-grams
3-grams
4-grams
N.Bayes
3 classes
Using harmony
87.8
88.4
88.4
87.8
9 classes
Using melody
54.7
48.8
47.7
58.7
Classifier ensembles
Best single
3 classes
88.4
(melody 2-grams)
Linear
Quadratic
90.1
90.1
Hierarchical classification
9 classes
60.5
62.8
63.4
1st level (3 classes): harmony 2-grams
2nd level (3x3 classes):
academic – melody NB
jazz – melody NB
popular – melody 4-grams
SSPR 2008
14
Conclusions
►
Harmony and melody are suitable features for music
genre classification
►
Harmony is better for classifying broad musical domains,
while melody is better for distinguishing sub-genres
►
Misclassifications occur more frequently within broad
domains
►
Hierarchical classification and classifier ensembles
outperformed the best single classifiers
SSPR 2008
15
Stochastic Text Models for Music Categorization
Carlos Pérez-Sancho, José M. Iñesta, David Rizo
Pattern Recognition and Artificial Intelligence group
Department of Software and Computing Systems
University of Alicante, Spain