- 1.Outline Problem FormulationMotivationProposed Method
Experimental Results Future WorkMusic Genre Classication Using
ExplicitSemantic Analysis and Sparsity-Eager SupportVector Machines
Kamelia AryafarDrexel University Computer Science Department
February 18, 2012 Kamelia AryafarMusic Genre Classication Using
Explicit Semantic Analysis
2. OutlineProblem Formulation Motivation Proposed
MethodExperimental ResultsFuture Work1 Problem Formulation2
MotivationChallengesRelated Work3 Proposed MethodFeature
SelectionFractional TF-IDFSparsity-Eager SVM Genre Classication4
Experimental ResultsBenchmark Data setResults5 Future WorkKamelia
AryafarMusic Genre Classication Using Explicit Semantic Analysis 3.
Outline Problem FormulationMotivationChallengesProposed Method
Related Work Experimental Results Future WorkMotivation Many
systems are exposed to high-dimensional data, e.g. images, image
sequences and even scalar signals. The high dimensional data could
be also multimodal. Kamelia AryafarMusic Genre Classication Using
Explicit Semantic Analysis 4. Outline Problem
FormulationMotivationChallengesProposed Method Related Work
Experimental Results Future WorkMotivation(Multimodal Mixture)
(Source I) (Source II) Kamelia AryafarMusic Genre Classication
Using Explicit Semantic Analysis 5. Outline Problem
FormulationMotivation ChallengesProposed MethodRelated Work
Experimental Results Future WorkBSS IllustrationArticial gaussian
mixture of two audio sources:(Violin mixture)(I)(II) Kamelia
Aryafar Music Genre Classication Using Explicit Semantic Analysis
6. Outline Problem FormulationMotivationChallengesProposed Method
Related Work Experimental Results Future WorkMotivationThe problem
of genre classication: (Violin playing) Kamelia AryafarMusic Genre
Classication Using Explicit Semantic Analysis 7. Outline Problem
FormulationMotivationChallengesProposed Method Related Work
Experimental Results Future WorkMotivationThe problem of genre
classication: (Violin playing)Genre Label: Classic Music/Violin
Kamelia AryafarMusic Genre Classication Using Explicit Semantic
Analysis 8. Outline Problem FormulationMotivationChallengesProposed
Method Related Work Experimental Results Future WorkMusic Genre
ClassicationGoalMusic genre classication is the problem of
categorization of apiece of music into its corresponding
categorical labels. Thegoal of automatic music genre classication
is to estimategenre labels for test audio sequences in large data
sets. Kamelia AryafarMusic Genre Classication Using Explicit
Semantic Analysis 9. Outline Problem
FormulationMotivationChallengesProposed Method Related Work
Experimental Results Future WorkMusic Genre ClassicationGoalMusic
genre classication is the problem of categorization of apiece of
music into its corresponding categorical labels. Thegoal of
automatic music genre classication is to estimategenre labels for
test audio sequences in large data sets.MotivationExponential
growth in available music data setsCost reductionExtension to
similar tasks Kamelia AryafarMusic Genre Classication Using
Explicit Semantic Analysis 10. Outline Problem
FormulationMotivationChallengesProposed Method Related Work
Experimental Results Future WorkChallenges Kamelia AryafarMusic
Genre Classication Using Explicit Semantic Analysis 11.
OutlineProblem Formulation MotivationChallenges Proposed Method
Related WorkExperimental ResultsFuture WorkChallenges The robust
representation of audio signals in terms of low-level features or
high-level audio keywords The construction of an automatic learning
schema to classify these representative features into music
genres.Kamelia AryafarMusic Genre Classication Using Explicit
Semantic Analysis 12. Outline Problem
FormulationMotivationChallengesProposed Method Related Work
Experimental Results Future WorkProposed MethodKamelia Aryafar
Music Genre Classication Using Explicit Semantic Analysis 13.
Outline Problem FormulationMotivationChallengesProposed Method
Related Work Experimental Results Future WorkProposed
MethodAbstract layer to represent features in terms of
conceptsInvariant to feature selection Kamelia AryafarMusic Genre
Classication Using Explicit Semantic Analysis 14. OutlineProblem
Formulation MotivationChallenges Proposed Method Related
WorkExperimental ResultsFuture WorkTF-IDF RepresentationGoalCreate
a high-level abstraction of low-level audio features(codewords of
MFCCs) to enhance music genre classication.Kamelia AryafarMusic
Genre Classication Using Explicit Semantic Analysis 15. Outline
Problem FormulationMotivationChallengesProposed Method Related Work
Experimental Results Future WorkTF-IDF RepresentationGoalCreate a
high-level abstraction of low-level audio features(codewords of
MFCCs) to enhance music genre classication.ESA ModelExplicit
semantic analysis (ESA) utilizes term-frequency (tf) andinverse
document frequency (idf) weighting schemata torepresent low-level
textual information in terms of concepts inhigher-dimensional
space. Kamelia AryafarMusic Genre Classication Using Explicit
Semantic Analysis 16. OutlineProblem Formulation
MotivationChallenges Proposed Method Related WorkExperimental
ResultsFuture WorkTF-IDF RepresentationEC,D [i, j] = tdf (Ci , j
).Kamelia AryafarMusic Genre Classication Using Explicit Semantic
Analysis 17. Outline Problem
FormulationMotivationChallengesProposed Method Related Work
Experimental Results Future WorkTF-IDF Representation EC,D [i, j] =
tdf (Ci , j ).TF-IDFThe relationship between a codeword and a
concept(document) pair will be captured through the so-called
tf-idfvalue of the word-concept pair. Kamelia AryafarMusic Genre
Classication Using Explicit Semantic Analysis 18. Outline Problem
FormulationFeature SelectionMotivationFractional TF-IDFProposed
MethodSparsity-Eager SVM Genre Classication Experimental Results
Future WorkMel Frequency Cepstral CoefcientsMFCCs represent
short-term power spectrum of sound and areknown to be effective for
music classication systems. Kamelia AryafarMusic Genre Classication
Using Explicit Semantic Analysis 19. Outline Problem
FormulationFeature SelectionMotivationFractional TF-IDFProposed
MethodSparsity-Eager SVM Genre Classication Experimental Results
Future WorkMel Frequency Cepstral CoefcientsMFCCs represent
short-term power spectrum of sound and areknown to be effective for
music classication systems.Pre-processingFor a large data set,
k-means clusteringof MFCCs creates the audio code-book,D = {1 ,
..., k }, using the cosinesimilarity distance measure to reduce
thecomplexity of the feature space. Kamelia AryafarMusic Genre
Classication Using Explicit Semantic Analysis 20. Outline Problem
FormulationFeature SelectionMotivationFractional TF-IDFProposed
MethodSparsity-Eager SVM Genre Classication Experimental Results
Future WorkFractional TF-IDF [2] Kamelia AryafarMusic Genre
Classication Using Explicit Semantic Analysis 21. Outline Problem
FormulationFeature SelectionMotivationFractional TF-IDFProposed
MethodSparsity-Eager SVM Genre Classication Experimental Results
Future WorkFractional TF-IDF [2] tdf (C, ) = tf (C, ) idfEC,D [i,
j] = tdf (Ci , j ) Kamelia AryafarMusic Genre Classication Using
Explicit Semantic Analysis 22. OutlineProblem Formulation Feature
Selection Motivation Fractional TF-IDF Proposed Method
Sparsity-Eager SVM Genre ClassicationExperimental ResultsFuture
WorkConcept-based Representation of Audio FeaturesKamelia
AryafarMusic Genre Classication Using Explicit Semantic Analysis
23. OutlineProblem Formulation Feature Selection Motivation
Fractional TF-IDF Proposed Method Sparsity-Eager SVM Genre
ClassicationExperimental ResultsFuture WorkTraining the ClassierESA
representation of the training setThe set E(T ) of (ESA-vector,
label) pairs will be provided as thetraining data to a supervised
classier algorithm.Kamelia AryafarMusic Genre Classication Using
Explicit Semantic Analysis 24. OutlineProblem Formulation Feature
Selection Motivation Fractional TF-IDF Proposed Method
Sparsity-Eager SVM Genre ClassicationExperimental ResultsFuture
WorkTraining the ClassierESA representation of the training setThe
set E(T ) of (ESA-vector, label) pairs will be provided as
thetraining data to a supervised classier algorithm.OutcomeThe set
of hyperplanes that dene the gaps between genres,are the outcome of
the training on E(T ).Kamelia AryafarMusic Genre Classication Using
Explicit Semantic Analysis 25. Outline Problem FormulationFeature
SelectionMotivationFractional TF-IDFProposed MethodSparsity-Eager
SVM Genre Classication Experimental Results Future WorkGenre
ClassicationClassier selectionSparsity-Eager support vector machine
( 1 -SVM) is used toassign samples to their genre categories.
Kamelia AryafarMusic Genre Classication Using Explicit Semantic
Analysis 26. OutlineProblem Formulation Feature Selection
Motivation Fractional TF-IDF Proposed Method Sparsity-Eager SVM
Genre ClassicationExperimental ResultsFuture WorkGenre
ClassicationClassier selectionSparsity-Eager support vector machine
( 1 -SVM) is used toassign samples to their genre categories. 1
-SVMIn contrast to the the original 2 -SVM, only a small subset of
thetraining examples contribute to the formation of the
nalclassier.Kamelia AryafarMusic Genre Classication Using Explicit
Semantic Analysis 27. OutlineProblem Formulation Feature Selection
Motivation Fractional TF-IDF Proposed Method Sparsity-Eager SVM
Genre ClassicationExperimental ResultsFuture WorkSparsity-Eager
SVM[1]ClassicationGiven a set of M training examples, we aim to nd
a samplesubset such that (i) subset is sufciently sparse, and (ii)
theclassier has a sufciently low empirical loss and
thereforesufciently large separating margin.Kamelia AryafarMusic
Genre Classication Using Explicit Semantic Analysis 28.
OutlineProblem Formulation Feature Selection Motivation Fractional
TF-IDF Proposed Method Sparsity-Eager SVM Genre
ClassicationExperimental ResultsFuture WorkSparsity-Eager
SVM[1]ClassicationGiven a set of M training examples, we aim to nd
a samplesubset such that (i) subset is sufciently sparse, and (ii)
theclassier has a sufciently low empirical loss and
thereforesufciently large separating margin.Why 1 -SVM(i) obtaining
higher generalization accuracy on new (test)examples, (ii)
increasing the robustness against overtting tothe training
examples, and (iii) providing scalability in terms ofthe
classication complexity.Kamelia AryafarMusic Genre Classication
Using Explicit Semantic Analysis 29. Outline Problem
FormulationMotivationBenchmark Data setProposed Method Results
Experimental Results Future WorkData set Description Data set:
Genre Samples We use the publicly alternative 145 available
benchmarkblues120 dataset for audio electronic113 classication
andfolk-country 222 clustering proposed by funk soul/R&B47
Homburg et al [3]. Thejazz319 dataset containspop 116 samples of
1886 songsrap/hip-hop300 obtained from the rock504 Garageband site.
Kamelia AryafarMusic Genre Classication Using Explicit Semantic
Analysis 30. OutlineProblem Formulation MotivationBenchmark Data
set Proposed Method ResultsExperimental ResultsFuture
WorkExperimental SetupParameters setupValidation method: 10-fold
cross validationPerformance measure: classication accuracy
rateSimilarity measure: cosine distanceKamelia AryafarMusic Genre
Classication Using Explicit Semantic Analysis 31. OutlineProblem
Formulation MotivationBenchmark Data set Proposed Method
ResultsExperimental ResultsFuture WorkExperimental SetupParameters
setupValidation method: 10-fold cross validationPerformance
measure: classication accuracy rateSimilarity measure: cosine
distanceComparative featuresAggregation of MFCC features
(AM)Temporal, spectral and phase (TSPS)Kamelia AryafarMusic Genre
Classication Using Explicit Semantic Analysis 32. Outline Problem
FormulationMotivation Benchmark Data setProposed MethodResults
Experimental Results Future WorkGenre Classication Accuracy Results
ESA Classier AMTSPS k = 1000 k= 5000Random 22.39 21.68 29.51 25.40
k-NN35.83 47.40 48.59 51.88 SVM 40.81 51.81 53.76
57.81ComparisonAggregation of MFCC features (AM) and temporal,
spectral andphase (TSPS) features are compared to the
ESArepresentation of MFCC features. Kamelia Aryafar Music Genre
Classication Using Explicit Semantic Analysis 33. OutlineProblem
Formulation MotivationBenchmark Data set Proposed Method
ResultsExperimental ResultsFuture WorkTrue Positive Accuracy
Rate50l1SVMlogregression45l2SVMl1regression40 classification
accuracy rate (%) per genre353025201510 5 0 12
345678AlternativeBlues Electronic FolkCountryJazz Pop Rock
Rap/Hiphop Figure: True positive genre classication rateKamelia
AryafarMusic Genre Classication Using Explicit Semantic Analysis
34. OutlineProblem Formulation MotivationBenchmark Data set
Proposed Method ResultsExperimental ResultsFuture WorkClassier
Convergence TimeFigure: Classier convergence timeKamelia
AryafarMusic Genre Classication Using Explicit Semantic Analysis
35. OutlineProblem Formulation MotivationBenchmark Data set
Proposed Method ResultsExperimental ResultsFuture WorkClassication
Accuracy vs. Training Samples Figure: Accuracy rate for different
samplesKamelia AryafarMusic Genre Classication Using Explicit
Semantic Analysis 36. OutlineProblem Formulation Motivation
Proposed MethodExperimental ResultsFuture WorkFuture Work MFCC
Representation CCA Space Audio Signals
ESA-Encoding(concepts)...CCALyrics DataTF-IDF TF
Representation(concepts) RepresentationKamelia AryafarMusic Genre
Classication Using Explicit Semantic Analysis 37. OutlineProblem
Formulation Motivation Proposed MethodExperimental ResultsFuture
WorkFuture Work...MFCC RepresentationCCA SpaceAudio Query
ESAENCODING ... Paired TextualData (Lyrics)Kamelia AryafarMusic
Genre Classication Using Explicit Semantic Analysis 38. Outline
Problem FormulationMotivationProposed Method Experimental Results
Future WorkQuestions?Thank you![1] Kamelia Aryafar, sina Jafarpour,
and Ali Shokoufandeh.Automatic musical genre classication using
sparsity-eager support vector machines.In NIME12, 2012.[2] Kamelia
Aryafar and Ali Shokoufandeh.Music genre classication using
explicit semantic analysis.In Proceedings of the 1st international
ACM workshop on Music information retrieval with user-centered
andmultimodal strategies, MIRUM 11, pages 3338, New York, NY, USA,
2011. ACM. [3] Helge Homburg, Ingo Mierswa, Bulent Moller,
Katharina Morik, and Michael Wurst.A benchmark dataset for audio
classication and clustering.In ISMIR, pages 528531,
2005.AcknowledmentsThis work was funded in part by Ofce of Naval
Research (ONR) grant N00014-04-1-0363 and United StatesNational
Science Foundation grant N0803670.Kamelia AryafarMusic Genre
Classication Using Explicit Semantic Analysis