Dissertationen zum Thema „Multi-class classifiers“

There are plenty of solutions for the task of multi-class recognition. Unfortunately, these solutions are not always unanimous. Most of them are based on empirical experiments while statistical data features consideration is often omitted. That’s why questions like when and which method should be used, what the reliability of any chosen method is for solving a multi-class recognition task arise. In this dissertation two-stage multi-class decision methods are analyzed. Pair-wise classifiers able to better exploit statistical data features are used in the first stage of such methods. In the second stage a particular fusion rule of the first stage results is used to fuse the first stage results in order to produce the final classification decision. Complexity issues of pair-wise classifiers, training data size and precision of method quality estimation are pointed out in the research. The precision of algorithm highly depends on the data and the number of experiments performed (data permutation, division into training and testing data). It is shown that the declared superiority of some known algorithms is not reliable due to low precision of estimation. A detailed comparison of well known multi-class classification methods is performed and a new pair-wise classifier fusion method based on similar method used in multi-class classifier fusion is presented. The recommendations for multi-class classification task designer are provided. Methods which allow reducing classification... [to full text]
Daugelio klasių atpažinimo uždaviniams spręsti yra sukurta aibė sprendimų ir ne visada vieningų rekomendacijų. Dauguma jų paremta empiriniais bandymais, retai atsižvelgiama į statistines duomenų savybes. Dėl to sprendžiant daugelio klasių klasifikavimo uždavinį kyla klausimų, kurį metodą ir kada geriausia naudoti, koks vieno ar kito metodo patikimumas. Disertacijoje nagrinėjami dviejų pakopų sprendimo priėmimo metodai, kai pirmame etape sudaromi klasifikatoriai poroms (angl. pair-wise), sugebantys geriau išnaudoti klasių tarpusavio statistines savybes, o kitame etape yra atliekamas klasifikatorių poroms rezultatų apjungimas. Tyrime ypatingas dėmesys yra skiriamas klasifikatorių poroms sudėtingumui, mokymo duomenų kiekiui bei algoritmų kokybės įvertinimo tikslumui. Tikslumas labai priklauso nuo duomenų bei atliktų eksperimentų kiekio (duomenų permaišymo klasėse, juos skirstant į mokymo ir testavimo). Parodyta, jog dėl žemo įvertinimo tikslumo kai kurių publikuotų algoritmų deklaruojamas pranašumas prieš žinomus algoritmus nėra patikimas. Darbe atliktas detalus žinomų metodų palyginimas bei pristatytas naujai sukurtas klasifikatorių poroms apjungimo algoritmas, kuris yra paremtas analogišku algoritmu daugelio klasių klasifikatorių rezultatų apjungimui. Pateiktos bendros rekomendacijos, kaip projektuotojui elgtis daugelio klasių atveju. Pasiūlyti metodai, leidžiantys sumažinti klasifikavimo klaidą atliekant klasifikatorių poroms apjungimo koregavimą, kad algoritmas nebūtų... [toliau žr. visą tekstą]

Indoor localisation techniques in multi-floor environments are emerging for location based service applications. Developing an accurate location determination and time-efficient technique is crucial for online location estimation of the multi-floor localisation system. The localisation accuracy and computational complexity of the localisation system mainly relies on the performance of the algorithms embedded with the system. Unfortunately, existing algorithms are either time-consuming or inaccurate for simultaneous determination of floor and horizontal locations in multi-floor environment. This thesis proposes an improved multi-floor localisation technique by integrating three important elements of the system; radio map fingerprint database optimisation, floor or vertical localisation, and horizontal localisation. The main focus of this work is to extend the kernel density approach and implement multi-class machine learning classifiers to improve the localisation accuracy and processing time of the each and overall elements of the proposed technique. For fingerprint database optimisation, novel access point (AP) selection algorithms which are based on variant AP selection are investigated to improve computational accuracy compared to existing AP selection algorithms such as Max-Mean and InfoGain. The variant AP selection is further improved by grouping AP based on signal distribution. In this work, two AP selection algorithms are proposed which are Max Kernel and Kernel Logistic Discriminant that implement the knowledge of kernel density estimate and logistic regression machine learning classification. For floor localisation, the strategy is based on developing the algorithm to determine the floor by utilising fingerprint clustering technique. The clustering method is based on simple signal strength clustering which sorts the signals of APs in each fingerprint according to the strongest value. Two new floor localisation algorithms namely Averaged Kernel Floor (AKF) and Kernel Logistic Floor (KLF) are studied. The former is based on modification of univariate kernel algorithm which is proposed for single-floor localisation, while the latter applies the theory kernel logistic regression which is similar to AP selection approach but for classification purpose. For horizontal localisation, different algorithm based on multi-class k-nearest neighbour ( NN) classifiers with optimisation parameter is presented. Unlike the classical kNN algorithm which is a regression type algorithm, the proposed localisation algorithms utilise machine learning classification for both linear and kernel types. The multi-class classification strategy is used to ensure quick estimation of the multi-class NN algorithms. The proposed algorithms are compared and analysed with existing algorithms to confirm reliability and robustness. Additionally, the algorithms are evaluated using six multi-floor and single-floor datasets to validate the proposed algorithms. In database optimisation, the proposed AP selection technique using Max Kernel could reduce as high as 77.8% APs compared to existing approaches while retaining similar accuracy as localisation algorithm utilising all APs in the database. In floor localisation, the proposed KLF algorithm at one time could demonstrate 93.4% correct determination of floor level based on the measured dataset. In horizontal localisation, the multi-class NN classifier algorithm could improve 19.3% of accuracy within fingerprint spacing of 2 meters compared to existing algorithms. All of the algorithms are later combined to provide device location estimation for multi-floor environment. Improvement of 43.5% of within 2 meters location accuracy and reduction of 15.2 times computational time are seen as compared to existing multi-floor localisation techniques by Gansemer and Marques. The improved accuracy is due to better performance of proposed floor and horizontal localisation algorithm while the computational time is reduced due to introduction of AP selection algorithm.

The work deals with an introduction to classification algorithms. It then divides classifiers into unary, binary and multi-class and describes the different types of classifiers. The work compares individual classifiers and their areas of use. For unary classifiers, practical examples and a list of used architectures are given in the work. The work contains a chapter focused on the comparison of the effects of hyper parameters on the quality of unary classification for individual architectures. Part of the submission is a practical example of reimplementation of the unary classifier.

Dans cet article, nous présentons une nouvelle approche de détection multi-classes basée sur un parcours hiérarchique de classifieurs appris simultanément. Pour plus de robustesse et de rapidité, nous proposons d’utiliser un arbre de classes d’objets. Notre modèle de détection est appris en combinant les contraintes de tri et de classification dans un seul problème d’optimisation. Notre formulation convexe permet d’utiliser un algorithme de recherche pour accélérer le temps d’exécution. Nous avons mené des évaluations de notre algorithme sur les benchmarks PASCAL VOC (2007 et 2010). Comparé à l’approche un-contre-tous, notre méthode améliore les performances pour 20 classes et gagne 10x en vitesse
Recent years have witnessed a competition in autonomous navigation for vehicles boosted by the advances in computer vision. The on-board cameras are capable of understanding the semantic content of the environment. A core component of this system is to localize and classify objects in urban scenes. There is a need to have multi-class object detection systems. Designing such an efficient system is a challenging and active research area. The algorithms can be found for applications in autonomous driving, object searches in images or video surveillance. The scale of object classes varies depending on the tasks. The datasets for object detection started with containing one class only e.g. the popular INRIA Person dataset. Nowadays, we witness an expansion of the datasets consisting of more training data or number of object classes. This thesis proposes a solution to efficiently learn a multi-class object detector. The task of such a system is to localize all instances of target object classes in an input image. We distinguish between three major efficiency criteria. First, the detection performance measures the accuracy of detection. Second, we strive low execution times during run-time. Third, we address the scalability of our novel detection framework. The two previous criteria should scale suitably with the number of input classes and the training algorithm has to take a reasonable amount of time when learning with these larger datasets. Although single-class object detection has seen a considerable improvement over the years, it still remains a challenge to create algorithms that work well with any number of classes. Most works on this subject extent these single-class detectors to work accordingly with multiple classes but remain hardly flexible to new object descriptors. Moreover, they do not consider all these three criteria at the same time. Others use a more traditional approach by iteratively executing a single-class detector for each target class which scales linearly in training time and run-time. To tackle the challenges, we present a novel framework where for an input patch during detection the closest class is ranked highest. Background labels are rejected as negative samples. The detection goal is to find the highest scoring class. To this end, we derive a convex problem formulation that combines ranking and classification constraints. The accuracy of the system is improved by hierarchically arranging the classes into a tree of classifiers. The leaf nodes represent the individual classes and the intermediate nodes called super-classes group recursively these classes together. The super-classes benefit from the shared knowledge of their descending classes. All these classifiers are learned in a joint optimization problem along with the previouslymentioned constraints. The increased number of classifiers are prohibitive to rapid execution times. The formulation of the detection goal naturally allows to use an adapted tree traversal algorithm to progressively search for the best class but reject early in the detection process the background samples and consequently reduce the system’s run-time. Our system balances between detection performance and speed-up. We further experimented with feature reduction to decrease the overhead of applying the high-level classifiers in the tree. The framework is transparent to the used object descriptor where we implemented the histogram of orientated gradients and deformable part model both introduced in [Felzenszwalb et al., 2010a]. The capabilities of our system are demonstrated on two challenging datasets containing different object categories not necessarily semantically related. We evaluate both the detection performance with different number of classes and the scalability with respect to run-time. Our experiments show that this framework fulfills the requirements of a multi-class object detector and highlights the advantages of structuring class-level knowledge

El texto completo de este trabajo no está disponible en el Repositorio Académico UPC por restricciones de la casa editorial donde ha sido publicado.
Protein fold recognition is an important task in the biological area. Different machine learning methods such as multiclass classifiers, one-vs-all and ensemble nested dichotomies were applied to this task and, in most of the cases, multiclass approaches were used. In this paper, we compare classifiers organized in tree structures to classify folds. We used a benchmark dataset containing 125 features to predict folds, comparing different supervised methods and achieving 54% of accuracy. An approach related to tree-structure of classifiers obtained better results in comparison with a hierarchical approach.
Revisión por pares

Associative Classification (AC) in data mining is a rule based approach that uses association rule techniques to construct accurate classification systems (classifiers). The majority of existing AC algorithms extract one class per rule and ignore other class labels even when they have large data representation. Thus, extending current AC algorithms to find and extract multi-label rules is promising research direction since new hidden knowledge is revealed for decision makers. Furthermore, the exponential growth of rules in AC has been investigated in this thesis aiming to minimise the number of candidate rules, and therefore reducing the classifier size so end-user can easily exploit and maintain it. Moreover, an investigation to both rule ranking and test data classification steps have been conducted in order to improve the performance of AC algorithms in regards to predictive accuracy. Overall, this thesis investigates different problems related to AC not limited to the ones listed above, and the results are new AC algorithms that devise single and multi-label rules from different applications data sets, together with comprehensive experimental results. To be exact, the first algorithm proposed named Multi-class Associative Classifier (MAC): This algorithm derives classifiers where each rule is connected with a single class from a training data set. MAC enhanced the rule discovery, rule ranking, rule filtering and classification of test data in AC. The second algorithm proposed is called Multi-label Classifier based Associative Classification (MCAC) that adds on MAC a novel rule discovery method which discovers multi-label rules from single label data without learning from parts of the training data set. These rules denote vital information ignored by most current AC algorithms which benefit both the end-user and the classifier's predictive accuracy. Lastly, the vital problem related to web threats called 'website phishing detection' was deeply investigated where a technical solution based on AC has been introduced in Chapter 6. Particularly, we were able to detect new type of knowledge and enhance the detection rate with respect to error rate using our proposed algorithms and against a large collected phishing data set. Thorough experimental tests utilising large numbers of University of California Irvine (UCI) data sets and a variety of real application data collections related to website classification and trainer timetabling problems reveal that MAC and MCAC generates better quality classifiers if compared with other AC and rule based algorithms with respect to various evaluation measures, i.e. error rate, Label-Weight, Any-Label, number of rules, etc. This is mainly due to the different improvements related to rule discovery, rule filtering, rule sorting, classification step, and more importantly the new type of knowledge associated with the proposed algorithms. Most chapters in this thesis have been disseminated or under review in journals and refereed conference proceedings.

Real life is full of multi-class decision tasks. In the Pattern Recognition field, several method- ologies have been proposed to deal with binary problems obtaining satisfying results in terms of performance. However, the extension of very powerful binary classifiers to the multi-class case is a complex task. The Error-Correcting Output Codes framework has demonstrated to be a very powerful tool to combine binary classifiers to tackle multi-class problems. However, most of the combinations of binary classifiers in the ECOC framework overlook the underlay- ing structure of the multi-class problem. In addition, is still unclear how the Error-Correction of an ECOC design is distributed among the different classes. In this dissertation, we are interested in tackling critic problems of the ECOC framework, such as the definition of the number of classifiers to tackle a multi-class problem, how to adapt the ECOC coding to multi-class data and how to distribute error-correction among different pairs of categories. In order to deal with this issues, this dissertation describes several proposals. 1) We define a new representation for ECOC coding matrices that expresses the pair-wise codeword separability and allows for a deeper understanding of how error-correction is distributed among classes. 2) We study the effect of using a logarithmic number of binary classifiers to treat the multi-class problem in order to obtain very efficient models. 3) In order to search for very compact ECOC coding matrices that take into account the distribution of multi-class data we use Genetic Algorithms that take into account the constraints of the ECOC framework. 4) We propose a discrete factorization algorithm that finds an ECOC configuration that allocates the error-correcting capabilities to those classes that are more prone to errors. The proposed methodologies are evaluated on different real and synthetic data sets: UCI Machine Learning Repository, handwriting symbols, traffic signs from a Mobile Mapping System, and Human Pose Recovery. The results of this thesis show that significant perfor- mance improvements are obtained on traditional coding ECOC designs when the proposed ECOC coding designs are taken into account.
En la vida cotidiana las tareas de decisión multi-clase surgen constantemente. En el campo de Reconocimiento de Patrones muchos métodos de clasificación binaria han sido propuestos obteniendo resultados altamente satisfactorios en términos de rendimiento. Sin embargo, la extensión de estos sofisticados clasificadores binarios al contexto multi-clase es una tarea compleja. En este ámbito, las estrategias de Códigos Correctores de Errores (CCEs) han demostrado ser una herramienta muy potente para tratar la combinación de clasificadores binarios. No obstante, la mayoría de arquitecturas de combinación de clasificadores binarios negligen la estructura del problema multi-clase. Sin embargo, el análisis de la distribución de corrección de errores entre clases es aún un problema abierto. En esta tesis doctoral, nos centramos en tratar problemas críticos de los códigos correctores de errores; la definición del número de clasificadores necesarios para tratar un problema multi-clase arbitrario; la adaptación de los problemas binarios al problema multi-clase y cómo distribuir la corrección de errores entre clases. Para dar respuesta a estas cuestiones, en esta tesis doctoral describimos varias propuestas. 1) Definimos una nueva representación para CCEs que expresa la separabilidad entre pares de códigos y nos permite una mejor comprensión de cómo se distribuye la corrección de errores entre distintas clases. 2) Estudiamos el efecto de usar un número logarítmico de clasificadores binarios para tratar el problema multi-clase con el objetivo de obtener modelos muy eficientes. 3) Con el objetivo de encontrar modelos muy eficientes que tienen en cuenta la estructura del problema multi-clase utilizamos algoritmos genéticos que tienen en cuenta las restricciones de los ECCs. 4) Pro- ponemos un algoritmo de factorización de matrices discreta que encuentra ECCs con una configuración que distribuye corrección de error a aquellas categorías que son más propensas a tener errores. Las metodologías propuestas son evaluadas en distintos problemas reales y sintéticos como por ejemplo: Repositorio UCI de Aprendizaje Automático, reconocimiento de símbolos escritos, clasificación de señales de tráfico y reconocimiento de la pose humana. Los resultados obtenidos en esta tesis muestran mejoras significativas en rendimiento comparados con los diseños tradiciones de ECCs cuando las distintas propuestas se tienen en cuenta.

Orientador: Siome Klein Goldenstein
Tese (doutorado) - Universidade Estadual de Campinas, Instituto da Computação
Made available in DSpace on 2018-08-12T17:37:15Z (GMT). No. of bitstreams: 1 Rocha_AndersondeRezende_D.pdf: 10303487 bytes, checksum: 243dccfe5255c828ce7ead27c27eb1cd (MD5) Previous issue date: 2009
Resumo: Neste trabalho de doutorado, propomos a utilizaçãoo de classificadores e técnicas de aprendizado de maquina para extrair informações relevantes de um conjunto de dados (e.g., imagens) para solução de alguns problemas em Processamento de Imagens e Visão Computacional. Os problemas de nosso interesse são: categorização de imagens em duas ou mais classes, detecçãao de mensagens escondidas, distinção entre imagens digitalmente adulteradas e imagens naturais, autenticação, multi-classificação, entre outros. Inicialmente, apresentamos uma revisão comparativa e crítica do estado da arte em análise forense de imagens e detecção de mensagens escondidas em imagens. Nosso objetivo é mostrar as potencialidades das técnicas existentes e, mais importante, apontar suas limitações. Com esse estudo, mostramos que boa parte dos problemas nessa área apontam para dois pontos em comum: a seleção de características e as técnicas de aprendizado a serem utilizadas. Nesse estudo, também discutimos questões legais associadas a análise forense de imagens como, por exemplo, o uso de fotografias digitais por criminosos. Em seguida, introduzimos uma técnica para análise forense de imagens testada no contexto de detecção de mensagens escondidas e de classificação geral de imagens em categorias como indoors, outdoors, geradas em computador e obras de arte. Ao estudarmos esse problema de multi-classificação, surgem algumas questões: como resolver um problema multi-classe de modo a poder combinar, por exemplo, caracteríisticas de classificação de imagens baseadas em cor, textura, forma e silhueta, sem nos preocuparmos demasiadamente em como normalizar o vetor-comum de caracteristicas gerado? Como utilizar diversos classificadores diferentes, cada um, especializado e melhor configurado para um conjunto de caracteristicas ou classes em confusão? Nesse sentido, apresentamos, uma tecnica para fusão de classificadores e caracteristicas no cenário multi-classe através da combinação de classificadores binários. Nós validamos nossa abordagem numa aplicação real para classificação automática de frutas e legumes. Finalmente, nos deparamos com mais um problema interessante: como tornar a utilização de poderosos classificadores binarios no contexto multi-classe mais eficiente e eficaz? Assim, introduzimos uma tecnica para combinação de classificadores binarios (chamados classificadores base) para a resolução de problemas no contexto geral de multi-classificação.
Abstract: In this work, we propose the use of classifiers and machine learning techniques to extract useful information from data sets (e.g., images) to solve important problems in Image Processing and Computer Vision. We are particularly interested in: two and multi-class image categorization, hidden messages detection, discrimination among natural and forged images, authentication, and multiclassification. To start with, we present a comparative survey of the state-of-the-art in digital image forensics as well as hidden messages detection. Our objective is to show the importance of the existing solutions and discuss their limitations. In this study, we show that most of these techniques strive to solve two common problems in Machine Learning: the feature selection and the classification techniques to be used. Furthermore, we discuss the legal and ethical aspects of image forensics analysis, such as, the use of digital images by criminals. We introduce a technique for image forensics analysis in the context of hidden messages detection and image classification in categories such as indoors, outdoors, computer generated, and art works. From this multi-class classification, we found some important questions: how to solve a multi-class problem in order to combine, for instance, several different features such as color, texture, shape, and silhouette without worrying about the pre-processing and normalization of the combined feature vector? How to take advantage of different classifiers, each one custom tailored to a specific set of classes in confusion? To cope with most of these problems, we present a feature and classifier fusion technique based on combinations of binary classifiers. We validate our solution with a real application for automatic produce classification. Finally, we address another interesting problem: how to combine powerful binary classifiers in the multi-class scenario more effectively? How to boost their efficiency? In this context, we present a solution that boosts the efficiency and effectiveness of multi-class from binary techniques.
Doutorado
Engenharia de Computação
Doutor em Ciência da Computação

碩士
臺灣大學
資訊工程學研究所
98
The classification problem appears in many applications such as document classification and web page search. Support vector machine(SVM) is one of the most popular tools used in classification task. One of the component in SVM is the kernel trick. We use kernels to map data into a higher dimentional space. And this technique is applied in non-linear SVMs. For large-scale sparce data, we use the linear kernel to deal with it. We call such SVM as the linear SVM. There are many kinds of SVMs in which different loss functions are applied. We call these SVMs as L1-SVM and L2-SVM in which L1-loss and L2-loss functions are used respectively. We can also apply SVMs to deal with multi-class classification with one-against-one or one-against-all approaches. In this thesis several models such as logistic regression, L1-SVM, L2-SVM, Crammer and Singer, and maximum entropy will be compared in the multi-class classification task.

Thesis (Ph.D.)--University of Nebraska-Lincoln, 2008.
Title from title screen (site viewed Mar. 10, 2009). PDF text: vi, 77 p. : ill. ; 634 K. UMI publication number: AAT 3336828. Includes bibliographical references. Also available in microfilm and microfiche formats.

碩士
淡江大學
資訊工程學系碩士在職專班
98
Regardless that the associative classification (AC) [1][2] method normally ranks the sequence according to the prescribed criteria, yet in terms of the problem of rule dependency that exists between rules, under the identical confidence value, support value and length criteria, the sequence by which the rules are executed can still impact the classification results. The core of the thesis, focusing on rule ranking problems, entails for more than adopting the Lazy[3] method as the general ranking principle for conducting document classification focusing on 100% confidence level, but also by pruning the classified documents to recalculate the confidence value ranking, together with using a multilevel class priority concept, to examine how it affects the classification performance. The TFIDF[4] weighing and the minimum classification criteria derived from the preliminary classification using the Naïve Bayes[5] classifier are used to define a single still-mode threshold value, and the Naïve Bayes classifier used to classify documents unclassifiable by the associative classification method, aiming to resolve the problem of lowering the classification precision rate due to the preset categories when using the associative classifiers.

碩士
國立臺灣大學
工業工程學研究所
95
Discriminant classifier is a type of supervised machine learning technique. There are two approaches to it. One is the Fisher’s discriminant; the other is the Minimum-Squared-Error (MSE) discriminant. The MSE discriminant is usually used to deal with two-class problems. The multi-class MSE approach extends the MSE discriminant to allow problems with more than two classes by providing a set of orthonormal class-label vectors through the Gram-Schmidt process. The resulting class-label vectors are made orthonormal so that the discriminants can be orthogonal as well. However, by giving different linearly independent vectors to the Gram-Schmidt process, the resulting class-label vectors will be different and so do the corresponding discriminants. That is, the solution of multi-class MSE is not unique and may not be the optimal. This research develops an iterative algorithm to obtain the class-label vectors and the discriminant loadings simultaneously while the objective is achieved. The objective is to make the discriminant scores as close to its corresponding class labels as possible. The iterative process is proven to be converged by the power method. The multi-class discriminants found through this iterative algorithm is called multi-class iterative MSE discriminants (IMSED). Through discriminant approaches, we will obtain the discriminant score for each instance. To allocate the instances to classes, there are mainly two types of classification rules. One is distance based; the other is probability based. Four classification rules will be discussed in this research and a probabilistic classification rule will be developed. Iris dataset will be used to illustrate the iterative algorithm and the classification rules. Finally, two real-world data sets with multiple classes are used to compare the IMSED classifier with the Fisher’s discriminant classifier and the multi-class MSE discriminant classifier.

碩士
國立臺灣大學
工業工程學研究所
94
In general, there are two kinds of linear classification methods: one is MSE, and the other is FLD. Because linear methods are not sufficient to analyze the data with nonlinear patterns, the nonlinear methods KMSE and KFD are hence developed from MSE and FLD, respectively. Both transform the instances from the original attribute space to the high-dimensional feature space and then linear methods are applied. The objective of FLD and KFD is to find the directions on which the projection of training instances can provide the maximal separability of classes. FLD and KFD are known to be inefficient for datasets with a large amount of attributes and instances, respectively. To improve the computing efficiency, we use MSE for linear classification problems. However, MSE, like SVM, can use only the one-against-one or the one-against-the-rest approach to solve the multi-class problems. Both are inefficient compared to FLD and KFD where only one model is built to discriminate multiple classes simultaneously. Thus, we develop the multi-class MSE with Sherman-Woodbury formula to improve the computation efficiency. It can deal with multiple classes simultaneously by a class-labeling scheme. The different class-labeling schemes are determined by the Gram-Schmidt process. The nonlinear application, multi-class KMSE, is also developed from the multi-class MSE. Then, a simulated example is used to show how the proposed method works and to visualize the meaning of the class-labeling scheme. Finally, two real-world datasets are used for comparing the proposed method with other conventional methods.

碩士
臺中技術學院
流通管理系碩士班
99
This paper proposes a novel classification model in dealing with multi-class problems when confronting large scales of features and instances. Our classification model comprises four stages: feature evaluation, feature selection, feature extraction and inductive learning. In the first stage, the investigation of data diversity enhances the classification effect of Information Gain. The second stage strengthens the relevance analysis by introducing correlation analysis and then serves a more reliable mechanism for feature selection. In the third stage, principal component analysis is applied to generate multivariate classifier. The final stage proceeds inductive learning. In order to verify our methods, five large datasets with class number of 4~7 and thousands of instances are used in our experiments. The assessments of accuracy, discrimination capability and performance are empolyed to evaluate our multivariate classifier as compared to C4.5, CART, SVM and NaiveBayes. In the experimental results, although our classifier performs the similar accuracy and discrimination capability with four conventional classifiers, model training time and classification efficiency is significantly improved.

碩士
銘傳大學
資訊管理學系碩士班
93
On the research of performance of automatic text classification, the number of term selection that influence the performance of text classification. There are many researches which done terms extraction in the past. But in the period of our research, we detected that in the text of terms with low weight which can’t increase the performance of text classification, on the contrary become noise to reduce the accuracy. In addition, on the research of text classification, there are many kinds of classifiers has been developed. The performance of different classifier gets different results. In the past, the research is focus on weather the data can be classified to the right class or not. And it is also have been composed many classifiers to a system. It depends on the property of data to choice different classifier to get better performance. Although it had been improved the performance, but it always only defined a data to a single class to cause error. Furthermore, the efficiency of classification have been influence by composed too much classifiers. The data will be processed by all kinds of classifiers, and then choice which one is better. Therefore, we filter the terms which are not important between classes, and the same time we filter comparative noise between classes. Moreover we design a process of a single classifier model which can deal with multi-class data to solve the error of only defined the data into a single class. In the experimental results, we can improve the accuracy of classification from 60%-75% up to 95%-100%.

Distributed optimization over multi-agent networks has become an increasingly popular research topic as it incorporates many applications from various areas such as consensus optimization, distributed control, network resource allocation, large scale machine learning, etc. Parallel distributed solution algorithms are highly desirable as they are more scalable, more robust against agent failure, align more naturally with either underlying agent network topology or big-data parallel computing framework. In this dissertation, we consider a multi-agent optimization formulation where the global objective function is the summation of individual local objective functions with respect to local agents' decision variables of different dimensions, and the constraints include both local private constraints and shared coupling constraints. Employing and extending tools from the monotone operator theory (including resolvent operator, operator splitting, etc.) and fixed point iteration of nonexpansive, averaged operators, a series of distributed solution approaches are proposed, which are all iterative algorithms that rely on parallel agent level local updates and inter-agent coordination. Some of the algorithms require synchronizations across all agents for information exchange during each iteration while others allow asynchrony and delays. The algorithms' convergence to an optimal solution if one exists are established by first characterizing them as fixed point iterations of certain averaged operators under certain carefully designed norms, then showing that the fixed point sets of these averaged operators are exactly the optimal solution set of the original multi-agent optimization problem. The effectiveness and performances of the proposed algorithms are demonstrated and compared through several numerical examples.

碩士
國立中正大學
資訊管理所
95
Regarding the non-skewed distribution, to utilize the existing data mining classification to construct the prediction model can reach a certain level of prediction accuracy. However, in the real data mining case, the dataset distribution is always skewed distribution. In clinical case, because the number of healthy people is more than the number of unhealthy people, the collected data would be congenital skewed distribution. If we utilize those dataset with skewed distribution to construct the prediction model, the prediction deviation should be a big problem. There are three existing solutions for skewed distribution – Under-sampling, Over-sampling, and Multi-classifier Committee Approach. This research will utilize Over-sampling and Multi-classifier Committee Approach for skewed distribution and improve them. The research objective is to raise the prediction accuracy of the minor part of the dataset. The case study is the disease of benign prostate hypertrophy and cancer of prostate. And this research will use those data to test the classification efficiency of my algorithm.

Presently, internal combustion engines provide power to move the majority of vehicles on the roadway. While battery-powered electric vehicles provide an alternative, their widespread acceptance is hindered by range anxiety and longer charging/refueling times. Dynamic wireless power transfer (DWPT) has been proposed as a means to reduce both range anxiety and charging/refueling times. In DWPT, power is provided to a vehicle in motion using electromagnetic fields transmitted by a transmitter embedded within the roadway to a receiver at the underside of the vehicle. For commercial vehicles, DWPT often requires transferring hundreds of kW through a relatively large airgap (> 20 cm). This requires a high-power DC-AC converter at the transmitting end and a DC-AC converter within the vehicle.

In this research, a focus is on the development of models that can be used to support the design of DWPT systems. These include finite element-based models of the transmitter/receiver that are used to predict power transfer, coil loss, and core loss in DWPT systems. The transmitter/receiver models are coupled to behavioral models of power electronic converters to predict converter efficiency, mass, and volume based upon switching frequency, transmitter/receiver currents, and source voltage. To date, these models have been used to explore alternative designs for a DWPT intended to power Class 8-9 vehicles on IN interstates. Specifically, the models have been embedded within a genetic algorithm-based multi-objective optimization in which the objectives include minimizing system mass and minimizing loss. Several designs from the optimization are evaluated to consider practicality of the proposed designs.