Maestría en Ingeniería · 2015
The classification problem in machine learning: an overview with study cases in emotion recognition and music-speech differentiation
This work addresses the well-known classification problem in machine learning -- The goal of this study is to approach the reader to the methodological aspects of the feature extraction, feature selection and classifier performance through simple and understandable theoretical aspects and two study cases -- Finally, a very good classification performance was obtained for the emotion recognition from speech
Full text 59 pages with text
Read the full thesis Record in the repository
Contents
- Introductionp. 6
- Justificationp. 7
- Objectivep. 7
- Outlinep. 8
- Feature extractionp. 9
- Databasesp. 9
- Features or measurementsp. 9
- Preprocessingp. 6
- Study casep. 10
- Multiresolution analysisp. 10
- Extracted featuresp. 11
- Feature selectionp. 13
- Class differentiation measurementsp. 13
- ROCp. 13
- Scatter matricesp. 13
- Classification performance measurementp. 14
- Confusion matrixp. 14
- Accuracyp. 14
- Sensitivity and specificityp. 14
- Precisionp. 14
- F1 scorep. 14
- Scalar selectionp. 14
- Feature vector selectionp. 14
- Study casep. 10
- Music/Speech datasetp. 15
- Berlin datasetp. 15
- Dimensionality reductionp. 19
- Principal component analysisp. 19
- Fisher’s linear discriminant analysisp. 19
- Study casep. 10
- Music/Speech databasep. 19
- Berlin databasep. 21
- Classifiersp. 23
- Data partitioning and trainingp. 23
- Linear classifiersp. 23
- LDAp. 23
- Perceptronp. 26
- Support Vector Machinep. 29
- Non-Linear classifiersp. 33
- Naive Bayesp. 33
- Decision treep. 35
- Artificial Neural Networkp. 37
- Kernelsp. 40
- Classification results for datasetsp. 48
- Emotion recognition from speechp. 49
- Backgroundp. 49
- Methodology and resultsp. 49
- Feature extractionp. 9
- Feature selection and classificationp. 50
- Conclusionsp. 55
- Referencesp. 56
- Figure 1 Steps for Classification taskp. 6
- Figure 2 Data treatment processp. 16
- Figure 3 Sequential forward floating selection of two features (Br.1)p. 16
- Figure 4 First two features of the composite feature ranking (Br.3)p. 17
- Figure 5 Energy of the signal and the zero crossing rate (Br.4)p. 18
- (M/S.P1)p. 20
- Figure 7 2D projection of the M/S database selected features (M/S.P2)p. 20
- (M/S.P3)p. 21
- Figure 9 PCA projection of a four feature sequential forward floating selection (Br.2)p. 22
- Br.3 and Br.4p. 48
- Br.1, Br.2, Br.3 and Br.4p. 29
- Br.2, Br.3 and Br.4p. 43
- Br.1, Br.2, Br.3 and Br.4p. 35
- Br.1, Br.2, Br.3 and Br.4p. 37
- Br.2, Br.3 and Br.4p. 43
- M/S.P2, M/S.P3, Br.1, Br.2, Br.3 and Br.4p. 43
- M/S.P2, M/S.P3, Br.1, Br.2, Br.3 and Br.4p. 45
- M/S.P3, Br.1, Br.2, Br.3 and Br.4p. 48
- featuresp. 8
- Figure 20 Confusion matrix for the ANN with exhaustive search selected featuresp. 53
- Figure 21 Confusion matrix for an ANN with the features selected by the genetic algorithmp. 54
- Table 1 Extracted features which are commonly used in signalsp. 12
- Table 2 Final features for the M/S databasep. 15
- Table 3 Features for each of the four Berlin database's selectionsp. 17
- Table 4 LDA accuracy for each datasetp. 24
- Table 5 Perceptron accuracy with each datasetp. 27
- Table 6 SVM accuracy in each of the datasetsp. 30
- Table 7 Better approximation for SVM for each dataset and the corresponding parameterp. 33
- Table 8 Naive Bayes accuracy for each datasetp. 33
- Table 9 Decision tree accuracy for each of the datasetsp. 36
- Table 10 ANN accuracy for each of the datasetsp. 38
- accuracyp. 14
- Table 12 Perceptron with RBF kernel accuracy for each datasetp. 41
- perceptron's accuracyp. 41
- Table 14 Perceptron with poly kernel accuracy for each datasetp. 43
- its accuracyp. 43
- Table 16 Accuracy in each dataset for a SVM with RBF kernelp. 46
- Table 17 Better parameters for each dataset and accuracy for the SVM with RBF kernelp. 46
- Table 18 Best accuracy for each dataset and its corresponding classifierp. 48
- Table 19 Dynamic featuresp. 50
- Table 20 Final 29 features selected with sequential floating forward feature selectionp. 51
- Table 21 Features selected with exhaustive searchp. 53
- Table 22 Features selected with the genetic algorithmp. 54
Keywords
Automatic speech recognitionElectromyographyEmotionsGenetic algorithmsMan-machine systemsNeural networks (Computer science)Pattern recognition systemsSpectrum analysisAnálisis discriminante linealHeurísticaTransformadas de Wavelet