This study concentrates on the comparison of the discriminant functions and the decision tree ind... more This study concentrates on the comparison of the discriminant functions and the decision tree induction techniques in antepartum fetal evaluation. These classification techniques are applied to antenatal fetal risk assessment problem and the performances, the computational complexities and the importance of each technique in terms of diagnostic clues are observed. The task is to investigate the Doppler ultrasound measurements of umbilical artery (UA) to relate the health conditions of fetuses using discriminant functions ...
2007 4th IEEE Workshop on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications, 2007
ABSTRACT Reliable arrhythmia classification from complex electrocardiogram (ECG) signals is one o... more ABSTRACT Reliable arrhythmia classification from complex electrocardiogram (ECG) signals is one of the most challenging pattern recognition problems. Several individual classifiers have been studied in the ECG domain. Also, parallel and serial classifier fusion systems have been proposed to increase the reliability. In this study, we are mainly interested in producing high confident arrhythmia classification results to be applicable in diagnostic decision support systems. We first experiment and compare two common techniques: support vector machines (SVM) and logistic regression (LR). Then, we propose a two- stage serial fusion classifier system based on SVM's rejection option. We relate the SVM's distance outputs to confidence measure and reject to classify ambiguous samples with first level SVM classifier. A non-symmetric thresholding scheme is applied: two different rejection distance thresholds have been defined for positive and negative ECG samples. The rejected samples have been forwarded to a second stage LR classifier. Finally we choose a way to combine the classifiers decisions to obtain a final decision rule. The experiments have been performed on UCI Arrhythmia Database.
The study discusses the nature of fuzziness in oxygen saturation (SaO 2 ) values taken by spectro... more The study discusses the nature of fuzziness in oxygen saturation (SaO 2 ) values taken by spectrophotometry measurements during intranatal fetal monitoring. The SaO 2 were taken from umbilical artery (UA) and umbilical vein (UV) with corresponding gestational week. Then, we employ a set of fuzzy rules relates SaO 2 values to output pH values in a neurofuzzy system. The output of the neurofuzzy system generates values for intranatal monitoring.
Istanbul Stock Exchange (ISE) stock market is not an efficient market. In this paper, we show how... more Istanbul Stock Exchange (ISE) stock market is not an efficient market. In this paper, we show how localized and global artificial neural network (ANN) models are used for risk estimation of asset returns. Mixture of Experts (MoE) and Recurrent Neural Networks (RNN) are more powerful to say that Efficient Market Hypothesis (EMH) is violated. ISE index XU100 is studied using daily data over a 14-year period using MoE and RNN neural networks and also Glosten-Jaganathan-Runkle (GJR) volatility models. The results suggest that localized neural approaches have the strength in modeling the risk in stock market time series data set of XU100.
Proceedings of 1993 International Conference on Neural Networks (IJCNN-93-Nagoya, Japan), 1993
Abstract The comparative performances of distributed and local neural networks for the speech rec... more Abstract The comparative performances of distributed and local neural networks for the speech recognition problem is investigated. Distributed networks' hidden units use the signoid nonlinearity with global response. We have used the backpropagation rule with three error measures: mean square error, cross entropy, and combinational performance. The hidden units of local networks respond only to inputs in a certain local region in the input space. We used k-nearest neighbor (kNN), Gaussian-based kNN, learning vector ...
This study proposes an intelligent data analysis approach to investigate and interpret the distin... more This study proposes an intelligent data analysis approach to investigate and interpret the distinctive factors of diabetes mellitus patients with and without ischemic (non-embolic type) stroke in a small population. The database consists of a total of 16 features collected from 44 diabetic patients. Features include age, gender, duration of diabetes, cholesterol, high density lipoprotein, triglyceride levels, neuropathy, nephropathy, retinopathy, peripheral vascular disease, myocardial infarction rate, glucose level, medication and blood pressure. Metric and non-metric features are distinguished. First, the mean and covariance of the data are estimated and the correlated components are observed. Second, major components are extracted by principal component analysis. Finally, as common examples of local and global classification approach, a k-nearest neighbor and a high-degree polynomial classifier such as multilayer perceptron are employed for classification with all the components ...
Proceedings of the IEEE 12th Signal Processing and Communications Applications Conference, 2004., 2004
In this study, the objective is to develop a new combined method for efficient compression of the... more In this study, the objective is to develop a new combined method for efficient compression of the classical 3D object mesh representation. This can be realized in two primary steps: mesh connectivity coding and data (geometry) compression. For realizing the first step, the algorithm of Isenburg (2002) has been employed. For the second step, vector quantization methods have been used
The representation of large document databases, consisting of Web pages, articles and book and ma... more The representation of large document databases, consisting of Web pages, articles and book and magazine titles, in terms of matrices for the purpose of text querying and retrieval simplifies and expedites the querying process. In the literature, dimensionality reduction techniques based on singular value decomposition and principal component analysis have been proposed to reduce the high computational complexity resulting from
In this study, we investigate several methods on the Interspeech 2013 Paralinguistic Challenge -S... more In this study, we investigate several methods on the Interspeech 2013 Paralinguistic Challenge -Social Signals Sub-Challenge dataset. The task of this sub-challenge is to detect laughter and fillers per frame. We apply Random Forests with varying number of trees and randomly selected features. We then proceed with minimum Redundancy Maximum Relevance (mRMR) ranking of features. We employ SVM with linear kernel to form a relative baseline for comparability to baseline provided in the challenge paper. The results indicate the relative superiority of Random Forests to SVMs in terms of subchallenge performance measure, namely UAAUC. We also observe that using mRMR based feature selection, it is possible to reduce the number of features to half with negligible loss of performance. Furthermore, the performance loss due to feature reduction is found to be less in Random Forests compared to SVMs. We also make use of neighboring frames to smooth the posteriors. On the overall, we attain an increase of 5.1% (absolute) in UAAUC in challenge test set.
In this study we present our system for INTERSPEECH 2014 Computational Paralinguistics Challenge ... more In this study we present our system for INTERSPEECH 2014 Computational Paralinguistics Challenge (ComParE 2014), Physical Load Sub-challenge (PLS). Our contribution is twofold. First, we propose using Low Level Descriptor (LLD) information as hints, so as to partition the feature space into meaningful subsets called views. We also show the virtue of commonly employed feature projections, such as Canonical Correlation Analysis (CCA) and Local Fisher Discriminant Analysis (LFDA) as ranking feature selectors. Results indicate the superiority of multi-view feature reduction approach to its single-view counterpart. Moreover, the discriminative projection matrices are observed to provide valuable information for feature selection, which generalize better than the projection itself. In our preliminary experiments we reached 75.35% Unweighted Average Recall (UAR) on PLS test set, using CCA based multi-view feature selection.
We propose a progressive mesh geometry coder, which expresses geometry information in terms of sp... more We propose a progressive mesh geometry coder, which expresses geometry information in terms of spectral coefficients obtained through a transformation and codes these coefficients using a hierarchical set partitioning algorithm. The spectral transformation used is the one proposed in where the spectral coefficients are obtained by projecting the mesh geometry onto an orthonormal basis determined by mesh topology. The set partitioning method that jointly codes the zeroes of these coefficients, treats the spectral coefficients for each of the three spatial coordinates with the right priority at all bit planes and realizes a truly embedded bitstream by implicit bit allocation. The experiments on common irregular meshes reveal that the distortion-rate performance of our coder is significantly superior to that of the spectral coder of .
In some machine learning problems, the dataset has multiple views which may be obtained using dif... more In some machine learning problems, the dataset has multiple views which may be obtained using different sensors or applying different sampling techniques. These views may have sufficient or partial information about the target concept. In this paper, a method that we called parallel interacting multiview learning (PIML) is proposed in which the views interact during the training process using the
Artificial Intelligence and Applications / 718: Modelling, Identification, and Control, 2011
Page 1. ISTANBUL STOCK EXCHANGE (ISE) 100 PREDICTION USING SUPPORT VECTOR PREDICTORS (SVP) M.Serd... more Page 1. ISTANBUL STOCK EXCHANGE (ISE) 100 PREDICTION USING SUPPORT VECTOR PREDICTORS (SVP) M.Serdar Yümlü Department of Computer Engineering Boğaziçi University 34342 Bebek, Istanbul, Turkey [email protected] ...
Proceedings of the 8th annual conference on Genetic and evolutionary computation - GECCO '06, 2006
Genetic Algorithms have widely been used for solving optimization problems in stationary environm... more Genetic Algorithms have widely been used for solving optimization problems in stationary environments. In recent years, there has been a growing interest for investigating and improving the performance of these algorithms in dynamic environments where the fitness landscape changes. In this study, we present an extensive comparison of several algorithms with different characteristics on a common platform by using the moving peaks benchmark and by varying problem parameters.
Location area (LA) management is a very important problem in mobile networks. In general, registr... more Location area (LA) management is a very important problem in mobile networks. In general, registration and paging costs are associated with tracking the current location of a mobile user. Considering minimizing the total of paging and registration costs as the main objective, the aim is to provide corresponding cell-to-switch and cell-to-LA assignments. This paper compares three well-known evolutionary algorithms to
This study presents parallel implementation of a vector quantization (VQ) based text-independent ... more This study presents parallel implementation of a vector quantization (VQ) based text-independent speaker identification system that uses Melfrequency cepstrum coefficients (MFCC) for feature extraction, Linde-Buzo-Gray (LBG) VQ algorithm for pattern matching ...
This study concentrates on the comparison of the discriminant functions and the decision tree ind... more This study concentrates on the comparison of the discriminant functions and the decision tree induction techniques in antepartum fetal evaluation. These classification techniques are applied to antenatal fetal risk assessment problem and the performances, the computational complexities and the importance of each technique in terms of diagnostic clues are observed. The task is to investigate the Doppler ultrasound measurements of umbilical artery (UA) to relate the health conditions of fetuses using discriminant functions ...
2007 4th IEEE Workshop on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications, 2007
ABSTRACT Reliable arrhythmia classification from complex electrocardiogram (ECG) signals is one o... more ABSTRACT Reliable arrhythmia classification from complex electrocardiogram (ECG) signals is one of the most challenging pattern recognition problems. Several individual classifiers have been studied in the ECG domain. Also, parallel and serial classifier fusion systems have been proposed to increase the reliability. In this study, we are mainly interested in producing high confident arrhythmia classification results to be applicable in diagnostic decision support systems. We first experiment and compare two common techniques: support vector machines (SVM) and logistic regression (LR). Then, we propose a two- stage serial fusion classifier system based on SVM's rejection option. We relate the SVM's distance outputs to confidence measure and reject to classify ambiguous samples with first level SVM classifier. A non-symmetric thresholding scheme is applied: two different rejection distance thresholds have been defined for positive and negative ECG samples. The rejected samples have been forwarded to a second stage LR classifier. Finally we choose a way to combine the classifiers decisions to obtain a final decision rule. The experiments have been performed on UCI Arrhythmia Database.
The study discusses the nature of fuzziness in oxygen saturation (SaO 2 ) values taken by spectro... more The study discusses the nature of fuzziness in oxygen saturation (SaO 2 ) values taken by spectrophotometry measurements during intranatal fetal monitoring. The SaO 2 were taken from umbilical artery (UA) and umbilical vein (UV) with corresponding gestational week. Then, we employ a set of fuzzy rules relates SaO 2 values to output pH values in a neurofuzzy system. The output of the neurofuzzy system generates values for intranatal monitoring.
Istanbul Stock Exchange (ISE) stock market is not an efficient market. In this paper, we show how... more Istanbul Stock Exchange (ISE) stock market is not an efficient market. In this paper, we show how localized and global artificial neural network (ANN) models are used for risk estimation of asset returns. Mixture of Experts (MoE) and Recurrent Neural Networks (RNN) are more powerful to say that Efficient Market Hypothesis (EMH) is violated. ISE index XU100 is studied using daily data over a 14-year period using MoE and RNN neural networks and also Glosten-Jaganathan-Runkle (GJR) volatility models. The results suggest that localized neural approaches have the strength in modeling the risk in stock market time series data set of XU100.
Proceedings of 1993 International Conference on Neural Networks (IJCNN-93-Nagoya, Japan), 1993
Abstract The comparative performances of distributed and local neural networks for the speech rec... more Abstract The comparative performances of distributed and local neural networks for the speech recognition problem is investigated. Distributed networks' hidden units use the signoid nonlinearity with global response. We have used the backpropagation rule with three error measures: mean square error, cross entropy, and combinational performance. The hidden units of local networks respond only to inputs in a certain local region in the input space. We used k-nearest neighbor (kNN), Gaussian-based kNN, learning vector ...
This study proposes an intelligent data analysis approach to investigate and interpret the distin... more This study proposes an intelligent data analysis approach to investigate and interpret the distinctive factors of diabetes mellitus patients with and without ischemic (non-embolic type) stroke in a small population. The database consists of a total of 16 features collected from 44 diabetic patients. Features include age, gender, duration of diabetes, cholesterol, high density lipoprotein, triglyceride levels, neuropathy, nephropathy, retinopathy, peripheral vascular disease, myocardial infarction rate, glucose level, medication and blood pressure. Metric and non-metric features are distinguished. First, the mean and covariance of the data are estimated and the correlated components are observed. Second, major components are extracted by principal component analysis. Finally, as common examples of local and global classification approach, a k-nearest neighbor and a high-degree polynomial classifier such as multilayer perceptron are employed for classification with all the components ...
Proceedings of the IEEE 12th Signal Processing and Communications Applications Conference, 2004., 2004
In this study, the objective is to develop a new combined method for efficient compression of the... more In this study, the objective is to develop a new combined method for efficient compression of the classical 3D object mesh representation. This can be realized in two primary steps: mesh connectivity coding and data (geometry) compression. For realizing the first step, the algorithm of Isenburg (2002) has been employed. For the second step, vector quantization methods have been used
The representation of large document databases, consisting of Web pages, articles and book and ma... more The representation of large document databases, consisting of Web pages, articles and book and magazine titles, in terms of matrices for the purpose of text querying and retrieval simplifies and expedites the querying process. In the literature, dimensionality reduction techniques based on singular value decomposition and principal component analysis have been proposed to reduce the high computational complexity resulting from
In this study, we investigate several methods on the Interspeech 2013 Paralinguistic Challenge -S... more In this study, we investigate several methods on the Interspeech 2013 Paralinguistic Challenge -Social Signals Sub-Challenge dataset. The task of this sub-challenge is to detect laughter and fillers per frame. We apply Random Forests with varying number of trees and randomly selected features. We then proceed with minimum Redundancy Maximum Relevance (mRMR) ranking of features. We employ SVM with linear kernel to form a relative baseline for comparability to baseline provided in the challenge paper. The results indicate the relative superiority of Random Forests to SVMs in terms of subchallenge performance measure, namely UAAUC. We also observe that using mRMR based feature selection, it is possible to reduce the number of features to half with negligible loss of performance. Furthermore, the performance loss due to feature reduction is found to be less in Random Forests compared to SVMs. We also make use of neighboring frames to smooth the posteriors. On the overall, we attain an increase of 5.1% (absolute) in UAAUC in challenge test set.
In this study we present our system for INTERSPEECH 2014 Computational Paralinguistics Challenge ... more In this study we present our system for INTERSPEECH 2014 Computational Paralinguistics Challenge (ComParE 2014), Physical Load Sub-challenge (PLS). Our contribution is twofold. First, we propose using Low Level Descriptor (LLD) information as hints, so as to partition the feature space into meaningful subsets called views. We also show the virtue of commonly employed feature projections, such as Canonical Correlation Analysis (CCA) and Local Fisher Discriminant Analysis (LFDA) as ranking feature selectors. Results indicate the superiority of multi-view feature reduction approach to its single-view counterpart. Moreover, the discriminative projection matrices are observed to provide valuable information for feature selection, which generalize better than the projection itself. In our preliminary experiments we reached 75.35% Unweighted Average Recall (UAR) on PLS test set, using CCA based multi-view feature selection.
We propose a progressive mesh geometry coder, which expresses geometry information in terms of sp... more We propose a progressive mesh geometry coder, which expresses geometry information in terms of spectral coefficients obtained through a transformation and codes these coefficients using a hierarchical set partitioning algorithm. The spectral transformation used is the one proposed in where the spectral coefficients are obtained by projecting the mesh geometry onto an orthonormal basis determined by mesh topology. The set partitioning method that jointly codes the zeroes of these coefficients, treats the spectral coefficients for each of the three spatial coordinates with the right priority at all bit planes and realizes a truly embedded bitstream by implicit bit allocation. The experiments on common irregular meshes reveal that the distortion-rate performance of our coder is significantly superior to that of the spectral coder of .
In some machine learning problems, the dataset has multiple views which may be obtained using dif... more In some machine learning problems, the dataset has multiple views which may be obtained using different sensors or applying different sampling techniques. These views may have sufficient or partial information about the target concept. In this paper, a method that we called parallel interacting multiview learning (PIML) is proposed in which the views interact during the training process using the
Artificial Intelligence and Applications / 718: Modelling, Identification, and Control, 2011
Page 1. ISTANBUL STOCK EXCHANGE (ISE) 100 PREDICTION USING SUPPORT VECTOR PREDICTORS (SVP) M.Serd... more Page 1. ISTANBUL STOCK EXCHANGE (ISE) 100 PREDICTION USING SUPPORT VECTOR PREDICTORS (SVP) M.Serdar Yümlü Department of Computer Engineering Boğaziçi University 34342 Bebek, Istanbul, Turkey [email protected] ...
Proceedings of the 8th annual conference on Genetic and evolutionary computation - GECCO '06, 2006
Genetic Algorithms have widely been used for solving optimization problems in stationary environm... more Genetic Algorithms have widely been used for solving optimization problems in stationary environments. In recent years, there has been a growing interest for investigating and improving the performance of these algorithms in dynamic environments where the fitness landscape changes. In this study, we present an extensive comparison of several algorithms with different characteristics on a common platform by using the moving peaks benchmark and by varying problem parameters.
Location area (LA) management is a very important problem in mobile networks. In general, registr... more Location area (LA) management is a very important problem in mobile networks. In general, registration and paging costs are associated with tracking the current location of a mobile user. Considering minimizing the total of paging and registration costs as the main objective, the aim is to provide corresponding cell-to-switch and cell-to-LA assignments. This paper compares three well-known evolutionary algorithms to
This study presents parallel implementation of a vector quantization (VQ) based text-independent ... more This study presents parallel implementation of a vector quantization (VQ) based text-independent speaker identification system that uses Melfrequency cepstrum coefficients (MFCC) for feature extraction, Linde-Buzo-Gray (LBG) VQ algorithm for pattern matching ...
Uploads
Papers by F. Gurgen