Several methods to enhance the efficiency of classifier ensembles have been proposed in the literature, applicable to both homogeneous and heterogeneous ensemble structures, mainly related to the definition of the ensemble structure. Basically, the ensemble structure selection can be done in two different ways, statically and dynamically. Unlike static selection, which uses the same structure throughout the whole training/testing phases, dynamic selection defines the ensemble structure for each test instance. Different dynamic selection methods have been proposed in the literature, mainly for ensemble members and dataset features, but little effort has been made to propose dynamic selection methods for combination methods, also known as fusion methods. Aiming to contribute to this important field, this paper proposes a method for the dynamic selection of combination methods in classifier ensembles, called DFS-ML (Dynamic Fusion Selection Meta-Learning). The method aims to enhance the classification system accuracy by selecting the most appropriate combination method for each testing instance, based on a meta-learning method. In order to assess the feasibility of the DFS-ML method, an empirical analysis is conducted. In this analysis, the proposed method is compared with 11 static combinations using 6 different classifier pool sizes. The empirical analysis demonstrated that DFS-ML outperforms traditional static combination approaches in nearly all the analyzed scenarios. The proposed method is capable of forming more adaptive and robust classification systems.

Many feature selection strategies have been proposed lately, using different criteria to select the most relevant features. The use of dynamic selection of attributes, however, showed that using multiple criteria simultaneously to determine the best subset of attributes for similar testing instances can provide encouraging results. Although the use of dynamic selection has alleviated some of the limitations found in traditional (static) selection methods, some characteristics in its processing lead to limitations of this technique on complex classification problems. In this context, this paper proposes an extension in the processing of a dynamic feature selection method based on the Pareto Front (named PF-DFS, which stands for Pareto Front – Dynamic Feature Selection), aiming to reach the total dynamicity of the PF-DFS processing and, as a consequence, improving its efficiency. The results found on the empirical analyses employed in this paper are promising, demonstrating that the extension proposed in this paper are able to obtain results superior to the attribute selection methods used as a comparative basis, as well as when compared to the original data set.

Processing instances quickly and updating models when a concept drift is detected is crucial in the data stream context. However, this process becomes even more challenging in the semi-supervised scenario in which only a few labelled instances are available. A semi-supervised framework called Dynamic Data Stream Learner (DyDaSL) has been proposed in literature to address this issue. This framework generates a fixed size ensemble in the first iteration to classify the remaining data. This paper proposes two extensions for DyDaSL, aiming to optimise the ensemble generation process. In both extensions, the ensemble size starts with only one classifier and it is increased throughout the learning process. In the first extension, a new classifier is included after a stream chunk is presented, while the second extension includes a new classifier when a drift is detected. In order to assess the feasibility of the proposed methods, an empirical analysis is conducted. As a result, these two proposed methods outperformed the standard variation in 128 out of 180 cases (71.11%), showing their effectiveness in all evaluated metrics.

The use of dynamic selection techniques in classifier ensembles, for attributes or member selection, has appeared in several studies in the literature as a mechanism to increase accuracy of classifier ensembles. In order to present an efficient classification structure, in this paper, we propose an automatic decision process to select the best classification structure for a testing instance. In order to do this, the decision process is capable of defining if a testing instance should be classified by a dynamic classifier ensemble (result of the integrated dynamic technique) or by a single classifier. The empirical analysis presents a substantial decrease in the processing time without deteriorating the performance of the classifier ensembles of the proposed technique.

Like other neurodevelopmental disorders, ADHD (attention deficit/hyperactivity disorder) typically manifests in preschool-aged children and results in impairments in an individual's personal, social, academic, or occupational functioning. According to the Diagnostic and Statistical Manual of Mental Disorders-V, there is no biological characteristic that is fundamental to the definition of ADHD diagnosis. However, several studies have already proposed neurobiological basis that demonstrate abnormalities in brain perfusion in different areas of the brain of an ADHD patient, but without consensus. For this reason, the aim of this study is to evaluate images of the brain SPECT (Single Photon Emission Computed Tomography) to assess whether there are brain regions that are more promising than others for ADHD diagnosis. Among the different brain regions analyzed, the one that has achieved the best accuracy in classifying ADHD is the frontal cortex with an accuracy of 80% in predicting ADHD.

Machine Learning (ML) is a field focused on developing methods and algorithms that allow machines to learn from previous data and experiences. There are several ML techniques that can be broadly divided into three main approaches: unsupervised, semi-supervised and supervised. The semi-supervised learning has been proposed as an attempt to solve some weaknesses of both supervised and unsupervised techniques. There are several semi-supervised techniques and co-training is one of the most used methods. In its original form, the co-training algorithm selects instances with the best prediction rates without necessarily considering the class to which they belong, which can be a problem in domains where there are classes with low representativeness. This paper proposes the inclusion of a data stratification strategy in the Co-training algorithm in order to maintain the same representativeness of the classes initially labeled throughout the learning process. The Co-training algorithm proposed in this work (Co-FlexCon-CS) was adapted from the original Co-FlexCon-C method with the inclusion of a data stratification technique. This study evaluates all the variations of the Co-FlexCon-CS, using 30 datasets with different characteristics and four classifiers algorithms, comparing it to the original Co-training and the Co-FlexCon-C. Finally, a statistical analysis was performed using Friedman Test, and the critical difference diagrams were evaluated for all methods using a post-hoc Friedman Test (Nemenyi). The obtained results show that the proposed approach achieves better results than the standard Co-training, which does not use this strategy.

Object Detection (OD) is an important task in Computer Vision with many practical applications. For some use cases, OD must be done on videos, where the object of interest has a periodic motion. In this paper, we formalize the problem of periodic OD, which consists in improving the performance of an OD model in the specific case where the object of interest is repeating similar spatio-temporal trajectories with respect to the video frames. The proposed approach is based on training a Gaussian Process to model the periodic motion, and use it to filter out the erroneous predictions of the OD model. By simulating various OD models and periodic trajectories, we demonstrate that this filtering approach, which is entirely data-driven, improves the detection performance by a large margin.

In architectural design of embedded systems, machine learning (ML) has become a promising solution to provide robustness to the design space exploration (DSE) of large hardware designs. However, given the large diversity of embedded applications, a main challenge in the design of a high-accuracy predictor is to select one ML algorithm to encompass a wide range of applications. In this context, regression ensemble is a promising solution since it can use multiple models and combine their predictions. In this work we employ the use of ensemble methods to predict performance when running different applications in different heterogeneous designs composed of a general purpose processor (GPPs) and a reconfigurable accelerator (RA). In our investigation, we evaluate three ensemble methods, Random Forest, AdaBoosting and Gradient Boosting. So, we compare them to the most used regression algorithms found in literature to perform DSE of computer architectures. Results show an error prediction rate below 2% on average when using ensemble methods and a throughput of more than 5,000 predictions per second when using Gradient Boosting.

The aim of this paper is to investigate the use of Dynamic Feature Selection in classifier ensembles, aiming at improving its diversity and the performance of these systems. In order to do that, four dynamic feature selection techniques are proposed and investigated in this paper. Therefore, an empirical analysis will be conducted, using different selection rates and ensemble sizes. Additionally, a comparative analysis will be conducted, with existing ensemble generation methods.

In real world classification problems, the amount of labelled data is usually limited (very hard or expensive to manually label the instances). However, a natural limitation of a classification algorithm is that it needs to have a set of labelled instances with a reasonable size in order to achieve a reasonable performance. Therefore, one solution to smooth out this problem is the use of semi-supervised learning. Several semi-supervised approaches (e.g. self training) have been proposed in the literature, aiming to use only a few labelled instances, to train a classifier, and to apply a labelling process in which a high number of unlabelled instances is labelled and included in the labelled set. However, this approach can include unreliable instances to the labelled set, impairing the performance of a semi-supervised method. In other words, the selection criterion to include newly labelled instances in the labelled set as well as the labelling step have an important effect in the performance of a semi-supervised method. In this paper, we propose two new approaches for automatic labelling in semi-supervised methods based on the prediction agreement of a pool of classifier as selection criterion. In addition, we compare them to the standard self-training method, and one variation of it called Flexible Confidence Classifier as baselines. In general, both methods obtained significantly better predictive results than the other two methods over 40 classification datasets.

Semi-supervised learning (SSL) is a paradigm that has been continuously used in data classification tasks in datasets that do not have enough labeled instances to train a supervised model with a minimum acceptable accuracy. In this context, data stream classification in dynamic environments appears as a natural application for this approach, because changes in data distribution contribute to decrease the performance of the classification algorithms. In this paper, we have proposed a framework, refered to as Dynamic Data Stream Learning (DyDaSL), that implements an auto-adaptive classifier ensemble - which is able to evaluate and replace classifiers with decreasing performance. This platform uses the FlexCon-C method, which is a variant of the self-training SSL algorithm that adjusts a confidence threshold dynamically, in each iteration, to define which instances will be labeled. Experimental tests on synthetic and real datasets show that the proposed approach obtains better results than traditional approaches using four evaluation metrics: accuracy, F-score, precision, and recall.

It has become increasingly common in academic and industrial environments the necessity to process huge amounts of seismic signals. Several researchers have been seeking for ways to improve and optimize the processing of these enormous amounts of data that are related to routine demands of geophysicists. One of these demands is the classification of distinct seismic regions captured by the same seismograph, a task that could take up to months of manual data processing. In this paper, we propose the usage of machine learning techniques to the task of the classification of seismic regions, in order to achieve accurate results with better performance and speed. The algorithms K-NN, MLP, Naive Bayes and Decision Tree were used for tests as base classifiers and also combined on ensemble methods. We also employed Deep Learning techniques, namely, a pure RNN network, and a variation of RNN called LSTM. The best results were achieved when using heterogeneous classifiers, showing accuracy rates of up to 98.52%. The results show that one can build an efficient seismic region classification system even when few classified data are already available for a specific seismograph setting.

Data classification tasks have been used in a wide range of problems in the Machine Learning field and the different learning paradigms (supervised, unsupervised or semi-supervised) define the task a computer can learn from a set of labeled and/or unlabeled data. This paper presents a study in the semi-supervised learning paradigm and proposes changes on the co-training algorithm in order to propose a confidence value procedure to include new instances in the labeled dataset. In order to evaluate the proposed method, an empirical analysis with 30 datasets has been conducted, with different characteristics, that were set up with different percentages of initially labeled instances. Each dataset was trained using four different classification algorithms (Naive Bayes, Decision tree, Ripper and k-NN) as basis for the co-training training procedure. The obtained results are promising and they indicate that, in most cases, the proposed method performs better than the co-training method originally proposed in the literature.

Fuzzy C-Means (FCM) is the most commonly used and discussed fuzzy clustering algorithm in the literature. Nevertheless, it is well known that the performance of FCM is strongly affected by the selection of the initial cluster centers. In other words, the selection of a good set of initial cluster c enters plays an important role in the performance of this algorithm. The most common selection method is the trial-and-test random method, in which each execution is performed with different initial centers, randomly generated, resulting in different dataset partitions. This paper proposes two methods to obtain the initial cluster centers which are applied in FCM and its variants. The proposed methods are deterministic, since, for each data set and number of clusters, they will always provide the same cluster centers set. The main advantage of these methods is to provide high quality partitions faster than the original methods as well as other FCM and ckMeans-based algorithms with deterministic selection of cluster centers.

Although several automatic computer systems have been proposed to address facial expression recognition problems, the majority of them still fail to cope with some requirements of many practical application scenarios. In this paper, one of the most influential and common issues raised in practical application scenarios when applying automatic facial expression recognition system, head pose variation, is comprehensively explored and investigated. In order to do this, two novel texture feature representations are proposed for implementing multi-view facial expression recognition systems in practical environments. These representations combine the block-based techniques with Local Ternary Pattern-based features, providing a more informative and efficient feature representation of the facial images. In addition, an in-house multi-view facial expression database has been designed and collected to allow us to conduct a detailed research study of the effect of out-of-plane pose angles on the performance of a multi-view facial expression recognition system. Along with the proposed in-house dataset, the proposed system is tested on two well-known facial expression databases, CK+ and BU-3DFE datasets. The obtained results shows that the proposed system outperforms current state-of-the-art 2D facial expression systems in the presence of pose variations.

In the context of facial expression recognition (FER), this paper reviews the fundamental theories of emotions and further explains the key dimensions of a defined emotional space. The main contribution of this paper is to propose a set of novel categorization methods for facial expressions to be used in the design of an automatic FER system. This novel categorization enables the facial expression to be interpreted in a better way that and to be more effective in practical applications of automatic FER systems. In order to validate the feasibility of the proposed categorization methods, a set of experiments is reported which investigates and analyzes the influence that the novel categorization brings to a multi-view FER system.

This paper presents a study in the field of semi-supervised learning and, more specifically, it proposes changes in the self-training algorithm in order to apply a data stratification method in the labeling process of this algorithm. Therefore, this work proposes a method, called FlexCon-CS, whose objective is to apply data stratification in the inclusion of new instances in the training data set. In this sense, the representativeness and class distribution will be maintained throughout the labeling process, with the same proportions of the initially labeled dataset. In order to evaluate this proposal, we performed experiments on 27 databases with different data distribution features. Each dataset was trained with four different classification algorithms, Naive Bayes, Decision Tree, ripper, and K-Nearest Neighbor classifiers. Moreover, the Friedman statistical test was applied to provide a statistically significant analysis of the obtained results. Our findings indicate that, in most cases, the proposed methods perform better than the original self-training method.

A feature selection method has the objective of selecting the best feature subset that represents the entire dataset. The majority of these methods apply a static selection procedure, since it selects a feature subset and uses it throughout the classification. Recently, dynamic feature selection has been emerged as an efficient alternative for feature selection. Instead of selecting the feature subset for the entire dataset, a dynamic method selects the best feature subset for an individual instance or a group of instances and, in this sense, each instance or group will have its own feature subset. The use of feature selection methods helps to improve the accuracy in classification tasks, using either single classifiers or ensemble of classifiers. In the context of ensembles, a feature selection method selects the best feature subset for each individual classifier to be used in an ensemble. In this paper, we propose an investigation of integrating dynamic feature selection (DFS) in ensemble of classifiers. More specifically, the use of DFS methods in dynamic ensemble methods, in which an ensemble structure (individual classifiers) is selected for each testing instance. Our main objective is to promote dynamicity in ensembles of classifiers in order to obtain more robust ensembles. In order to accomplish this investigation, two well known dynamic ensemble methods are chosen to be analyzed. Our findings indicated real benefits when integrating the dynamic feature selection method with the dynamic ensemble selection methods, for the majority of cases.

One of the main issues of machine learning algorithms is the curse of dimensionality. With the fast growing of complex data in real world scenarios, the feature selection becomes a mandatory preprocessing step in any application to reduce both the complexity of the data and the computing time. Based on that, several works have been produced in order to develop efficient methods to perform this task. Most feature selection methods select the best attributes based on some specific criteria. Additionally, recent studies have successfully constructed models to select features considering the particularities of the data, assuming that similar samples should be treated separately. Although some advance has been made, a bad choice of one single criteria to evaluate the importance of the attributes and the arbitrary choice of the number of features made by the user can lead to a poor analysis. In order to overcome some of these issues, this work brings an improvement of a dynamic feature selection algorithm (DFS) by using the idea of pareto front multi-objective optimization, which allow us to both consider distinct perspectives of the features relevance and automatically set the number of attributes to select. We tested our approach using 15 artificial and real world data and results have shown that when compared to the original DFS method, the performance of the proposed method is remarkable superior. In fact, the results are very promising since the proposed method also achieved better performance than well-established dimensionality reduction methods and when using the original datasets, showing that the reduction of noisy and/or redundant attributes can have a positive effect in the performance of a classification task.

In the last decades, the use of ensembles of classifiers in different domains of application has received significant attention of the Machine Learning community. In a typical architecture of ensemble, a new input pattern is presented to all K components. The individual classifiers provide their output and send them to a combination method, which is responsible for providing the final output. Ensemble of classifiers can be applied to any classification problem, single-label or multi- label problems. Multi-label (ML) classification has been received much attention from various research domains, such as text categorization, bioinformatics, computer vision, among others. In this paper, we investigate the use of ensemble of classifiers when applied to multi-label problems, focusing on the use of diversity measures in these systems when applied to multi-label problems. In order to do this, the present work proposes an adapted approach for two well-known diversity measures, good and bad, for ensembles of multi-label classifiers. In order to assess the feasibility of the proposed diversity measures in ensemble of multi-label classifiers, an empirical analysis is conducted, with eight multi-label classification problems. Generally, our finds indicate that there is a strong correlation between performance of the ML ensembles with the proposed diversity measures, showing that these diversity measures can be used as an important tool in the design of efficient ensembles of multi-label classifiers.

Classifier ensembles have been widely studied in the literature as an attempt to increase the performance of individual classification structures. An important issue in the design of classifier ensemble is the definition of its structure. More specifically, the selection of the best individual classifiers and the combination method for an ensemble. Usually, an exhaustive test-and-trial process may be needed to define its structure. In parallel, new contributions of meta-learning have been presented as an efficient alternative to the automatic recommendation of classification algorithms. In this paper, we will apply meta-learning in the process of recommendation of the combination method of ensemble systems. The main goal of this paper is to provide one step towards the automatic design of classifier ensembles. In order to achieve this goal, three different representation approaches for the meta-learning recommendation process are proposed. In addition, an empirical analysis is performed, in which five meta-learners are used to evaluate the recommendation performance.

This paper consists of a study in the field of semi-supervised learning and implements changes on the self-training algorithm in order to propose a variation in the rate of inclusion of new observations in the labeled dataset. In order to achieve this goal, three methods (FlexCon-G, FlexCone FlexCon-C) are proposed, which differ in the way that they perform the calculation of a new value for the minimum confidence rate to include new patterns. In order to evaluate the proposed methods, we performed experimentations with 20 datasets with diversified characteristics. Each of them was setup with a different percentage of initially labeled patterns. Each dataset was trained using the Naive Bayes, decision tree and ripper classifiers. Moreover, Friedmann statistical test was applied to provide a statistically significant analysis. The obtained results indicate that the three proposed methods perform better than a self-training method in most cases, pointing to the FlexCon-C method as the most efficient of them.

Usually, the evaluation of the classifiers performance is not an easy task to be performed, mainly when we analyze different criteria (output parameters). In this evaluation process, we can use quantitative measures (accuracy, specificity, among others), however, when the output values are very close and we have several criteria, the results are difficult to be interpreted by users. This paper aims to propose a new linguistic model to evaluate the performance of several classifiers. It is based on the notion of gradual complex numbers (GCN), proposed in [18]. In this paper, we present the theoretical basis of GCNs for classifier evaluator and we assess the performance of the proposed model (GCN) through an empirical study. In addition, the performance of GCN is compared with that of fuzzy complex numbers[6], and it reveals gains.

In recent years, metaheuristic algorithms have been extensively used to solve global optimization problems. In general, they are used in complex problems where a good solution is needed, especially in cases with incomplete or imperfect information. Recently, the concept of efficiently combine meta-heuristics has emerged, in a field called hybridization of meta-heuristics. These hybrid systems have been successfully applied in traditional optimization problems. In this paper, a hybrid system, called MAMH (Multi-agent Metaheuristic Hybridization) is adapted to combine trajectory-based metaheuristics and to be applied in the design of ensemble systems. The main goal of this paper is to evaluate the use of hybrid trajectory-based meta-heuristics applied to the design of ensembles of classifiers. In order to validate the feasibility of using MAMH as ensemble generator, an empirical analysis will be conducted, in which a comparative analysis between MAMH and traditional trajectory-based metaheuristics will be performed. Our findings indicated a competitive performance of MAMH, with the best performance for the most important objective function.

Ensemble systems are classification structures that apply a two‐level decision‐making process, in which the first level produces the outputs of the individual classifiers and the second level produces the output of the combination method (final output). Although ensemble systems have been proven to be efficient for pattern recognition tasks, its efficient design is not an easy task. This article investigates the influence of two diversity measures when used explicitly to guide the design of ensemble systems. These diversity measures were proposed recently, and they proved to be very interesting for the diversity–accuracy dilemma. To perform this investigation, we will use two well‐known optimization techniques, genetic algorithms, and tabu search, in their mono‐objective and multiobjective versions. As objectives of the optimization techniques, we use error rate and two diversity measures as well as all possible combinations of these three objectives. In this article, we aim to analyze which set of objectives can generate more accurate ensembles. In addition, we aim to analyze whether or not the diversity measures (good and bad diversities) have a positive effect in the design of ensemble systems, mainly if they can replace the error rate as an optimization objective without incurring significant losses in the accuracy level of the generated ensembles.

Recently, the number of features in different problem domains has grown enormously. In order to select the best representation (attributes) for these problems, a deep knowledge of the problem domain is required. As this type of knowledge is not always possible, feature selection needs to be applied as an automatic selection process of the most relevant attributes in a dataset. In this paper, we propose a new dynamic feature selection technique using data clustering algorithms to select features in a dynamic way and the selected features will be used in classification methods. Our technique aims to select the best attributes for a group of instances rather than to the entire dataset, leading to a dynamic way to select attributes. We will also carry out an empirical analysis using well-known existing methods. Our findings indicated gains when comparing the proposed methods to the existing ones, for the majority of cases.

Biometric systems use personal biological or behavioural traits that can uniquely characterise an individual but this uniqueness property also becomes its potential weakness when the template characterising a biometric trait is stolen or compromised. To this end, we consider two strategies to improving biometric template protection and performance, namely, (1) using multiple privacy schemes and (2) using multiple matching algorithms. While multiple privacy schemes can improve the security of a biometric system by protecting its template; using multiple matching algorithms or similarly, multiple biometric traits along with their respective matching algorithms, can improve the system performance due to reduced intra-class variability. The above two strategies lead to a novel, ensemble system that is derived from multiple privacy schemes. Our findings suggest that, under the worst-case scenario evaluation where the key or keys protecting the template are stolen, multi-privacy protection scheme can outperform a single protection scheme as well as the baseline biometric system without template protection.

Researches with ensemble Systems have emerged as an attempt to obtain a computational system that works with classification tasks in an efficient way. The main goal of using ensemble systems is to improve the performance of a pattern recognition system in terms of better generalization and/or of clearer design. One of the main challenges in the design of a ensemble system is the definition of the system components. The choice of the ensemble members can become a very difficult task and, in some cases, it can lead to ensembles with no performance improvement. In order to avoid this situation, the idea of DES (Dynamic Ensemble Selection)-based method has emerged, in which the classifiers to compose the ensemble systems are chosen in a dynamic way. In this paper, we present an analysis of different diversity measures in two dynamic ensemble election methods. These two methods use accuracy and diversity as the main criteria to select classifiers dynamically. The goal of this paper is to investigate the influence of different diversity measure in the dynamic selection of classifiers.

Several clustering algorithms have been developed and applied to a great variety of problems in different fields. However, some of these algorithms have limitations. Bio-inspired algorithms have been applied to clustering problems aiming to overcome some of these limitations. In this paper, we apply the Coral Reefs Optimization (CRO) algorithm to clustering problems. The CRO algorithm has been originally proposed for classical optimization problems. In this paper, this algorithm will be adjusted to provide a good clustering partition for a dataset. In addition, we also propose three new modifications of this algorithm and an index to be used as objective function for the optimization techniques. In order to evaluate the effectiveness of the CRO algorithm and the proposed extensions when dealing with real data, we conduct a comparison analysis with another bio-inspired algorithm, a hybrid genetic algorithm proposed for solving clustering problems. In this analysis, two clustering validity measures are employed to measure the generated clusters by the bio-inspired algorithms. We also use two objective functions (TWCV and MX index) in the reproduction process of the analysed algorithms.

The main goal of using data with interval nature is to represent numeric information endowed with impreciseness, which are normally captured from measures of real world. However, in order to do this, it is necessary to adapt real-valued techniques to be applied on interval-based data. For interval-based clustering applications, for instance, it is necessary to propose an interval-based distance and also to adapt clustering algorithms to be used in this context. Therefore, in this paper, we aim to provide a platform for performing clustering applications using interval-based data, including distance measure, clustering algorithms, and validation indexes. In this case, we adapt an interval-based distance called dkm, and we propose two interval-based fuzzy clustering algorithms: Interval-based FcM and interval-based ckMeans, and three interval-based validation indexes. In order to validate the proposed interval-based framework, an empirical analysis was conducted using seven clustering datasets, three real and four synthetic interval datasets. The empirical analysis is based on an external cluster validity index, corrected rand, and six internal-based validation indexes, in which three of them can be used in their original proposal and three are proposed in this paper. The obtained results show the usefulness of the proposed interval-based framework for interval-based clustering problems.

The main advantage of using an interval-based distance for interval-based data lies on the fact that it preserves the underlying imprecision on intervals which is usually lost when real-valued distances are applied. One of the main problems when using interval-based distance in fuzzy clustering algorithms is the way to obtain the center of the groups. In this case, it is necessary to make adaptations in order to obtain those centers. Therefore, in this paper, we propose the use of the family of H-operator to proposed three approaches to transform the interval-based membership matrix into real-valued membership matrix and, as a consequence, to calculate the centers of the groups in interval-based fuzzy clustering algorithms. In this case, we will perform a comparative analysis using the three different approaches proposed in this paper, using seven interval-based datasets (four synthetic and three real datasets). As a result of this analysis, we will observe that the proposed approaches achieved better performance than all analyzed methods for interval-based methods.

This paper presents an experimental analysis of a revocable biometric verification problem using ensemble systems. Behavioural Biometric-based systems are a future emergent area on identification, verification and access control systems of users. However, there is still progress to be done in this field, specially related to system security and acceptable results for practical use. Cancellable Biometrics is a alternative solution to the security problem of biometric data. This technique consists of applying transformation functions to biometric data in order to protect the original characteristics of biometric template. In this case, if biometric template has compromised, a new representation of original biometric data can be generated. Although cancellable biometrics were proposed to solve privacy concerns, this concept raises new issues, becoming the authentication problem more complex and difficult to solve. Thus, more effective authentication structures are needed to perform these tasks. This work aims to investigate the use of ensemble systems in cancellable behavioural biometric system used by million people (touchscreen devices). Apart this, we also present an empirical analysis, comparing the ensemble structures with single classification algorithms.

In this paper, we investigate two important problems in multi-label classification algorithms, which are: the number of labeled instances and the high dimensionality of the labeled instances. In the literature, we can find several papers about multi-label classification problems, where an instance can be associated with more than one label simultaneously. One of the main issues with multi-label classification methods is that many of these require a high number of instances to be able to generalize in an efficient way. In order to solve this problem, we used semi-supervised learning, which combines labeled and unlabeled instances during the training process. In this sense, the semi-supervised learning may become an essential tool to define, efficiently, the process of automatic assignment of labels. Therefore, this paper presents four semi-supervised methods for the multi-label classification, focusing on the use of a confidence parameter in the process of automatic assignment of labels. In order to validate the feasibility of these methods, an empirical analysis will be conducted using high-dimensional datasets, aiming to evaluate the performance of such methods in different situations. In this case, we will apply a feature selection algorithm in order to reduce, in an efficient way, the number of features to be used by the classification methods.

In classification problems with hierarchical structures of labels, the target function must assign several labels that are hierarchically organized. The hierarchical structures of labels can be used either for single-label (one label per instance) or multi-label classification problems (more than one label per instance). In general, classification tasks are usually trained using a standard supervised learning procedure. However, the majority of classification methods require a large number of training instances to be able to generalize the mapping function, making predictions with high accuracy. In order to smooth out this problem, the idea of semi-supervised learning has emerged. It combines labelled and unlabelled data during the training phase. Some semi-supervised methods have been proposed for single-label classification methods. However, very little effort has been done in the context of multi-label hierarchical classification. This paper proposes the use of a semi-supervised learning method for the multi-label hierarchical problems. In order to validate the feasibility of these methods, an empirical analysis will be conducted, comparing the proposed methods with their corresponding supervised versions. The main aim of this analysis is to observe whether the semi-supervised methods proposed in this paper have similar performance to the corresponding supervised versions.

This paper investigates the influence of measures of good and bad diversity when used explicitly to guide the search of a genetic algorithm to design ensemble systems. We then analyze what the best set of objectives between classification error, good diversity and bad diversity as well as all combination of them. In this analysis, we make use of the NSGA II algorithm in order to generate ensemble systems, using k-NN as individual classifiers and majority vote as the combination method. The main goal of this investigation is to determine which set of objectives generates more accurate ensembles. In addition, we aim to analyze whether or not the diversity measures (good and bad diversity) have a positive effect in the construction of ensembles and if they can replace the classification error as optimization objective without causing losses in the accuracy level of the generated ensembles.

This paper presents an experimental analysis in a revocable biometric verification problem. Behavioural Biometric-based systems are the emergent area in future of the user identification, verification and access control systems. However, there is still much progress to be done in this field, specially related to system security and acceptable accuracy results for practical use. One alternative solution to the security problem in biometric data is a technique known as cancellable biometrics. This technique consists of applying a transformation on the biometric data in order to protect the original characteristics. This work aims to investigate the use of cancellable transformation functions in a behavioural biometric system used everyday for million of people (touchscreen devices). The main goal is to analyse the advantages and challenges that cancellable transformation functions may bring to behavioural biometric verification systems.

Currently, there is a concern about the security of biometric data in the identification systems, mainly due to the increase of fraudulent attacks in these systems. Therefore, in this paper, we propose a comparative analysis of traditional cryptographic algorithms and transformation functions to be used as biometric template protection methods in the identification systems. In this sense, we aim to contribute with new approaches to the area of information security systems. Our goal is to analyse the increase of the biometric dataset security as well as the performance of these protected dataset in the biometric-based identification systems. In this comparative analysis, we apply well-elaborated structures, called ensemble systems, as the pattern recognition structure. These systems are applied to unprotected and protected biometric dataset to measure the performance of the template protection methods and traditional cryptographic algorithms used in this work. As a result of this comparative analysis, we intend identify which methods have a good trade-off between security and accuracy.

In a decision making process, we are usually oriented to take into consideration all the relevant features (characteristics) involved in a specific problem. In Machine Learning, for instance, a decision is made through the use of a learning algorithm and the characterization process is represented by the corresponding datasets. In this context, classification algorithms can be applied, individually or through the use of ensemble systems (combination of classification methods), in the decision-making process. The concept of ensemble systems has emerged in the last decades as a strategy for combining independent classifiers (components), aiming to provide a decision that is potentially more effective than any single component. However, the design of the ensemble structure is not an easy task and it can have an important impact in the performance of these systems. In this paper, we investigate the use of meta-learning on the selection of the best configuration parameters (learning strategy, size and individual classifiers) for homogeneous structures of ensemble systems. The main aim of this analysis is to assess the effect of using meta-learning in the design of efficient and robust ensemble systems.

Biometric-based identification systems can offer several advantages over traditional forms of identity authentication. However, concerns have been raised about the privacy of the personal biometric data, since these systems need to ensure their integrity and public acceptance. In order to address these issues, the notion of cancellable biometrics was introduced. It describes biometric templates that can be cancelled and replaced, in case of being lost or stolen. However, this concept still raises new issues, since they make the authentication problem more complex and difficult to solve. Thus, more effective authentication structures are needed to perform these tasks. In this paper, we investigate the use of ensemble systems in cancellable biometrics, using online signature identification. In order to improve the effectiveness of the ensemble systems, we used genetic algorithms in the choice of an optimized set of weights that are used along with the output of the individual classifiers to define the final output of the system. In addition, we proposed the use of genetic algorithm in the procedure to create the cancellable biometric data, aiming to obtain more efficient cancellable data. The main of this paper is to provide more security in the biometric-based identification process.

In most traditional classification methods, each instance is associated with one single nominal target variable (single-label problems). However, there are also cases where an instance can be associated with more than one label simultaneously, referring to as multi-label classification problems. One of the main problems with classification methods is that many of these require a high number of instances to be able to generalize the mapping function, making predictions with high accuracy. In order to smooth out this problem, the idea of semi-supervised learning has emerged. It combines labeled and unlabelled data during the training phase. However, in semi-supervised learning, it is important to define an efficient process of assignments of instances. This paper proposes three semi-supervised methods for the multilabel classification, focusing on the use of a confidence parameter in the process of automatic assignment of labels. In order to validate the feasibility of these methods, an empirical analysis will be conducted, aiming to evaluate the performance of such methods in different situations, besides the use of different evaluation metrics on this performance.

Similarity and dissimilarity (distance) between objects is an important aspect that must be considered when clustering data. When clustering categorical data, for instance, these distance (similarity or dissimilarity) measures need to address properly the real particularities of categorical data. In this paper, we perform a comparative analysis with four different dissimilarity measures used as a distance metric for clustering categorical data. The first one is the Simple Matching Dissimilarity Measure (SMDM), which is one of the simplest and the most used metric for categorical attribute. The other two are context-based approaches (DIstance Learning in Categorical Attributes — DILCA and Domain Value Dissimilarity-DVD), and the last one is an extension of the SMDM, which is proposed in this paper. All four dissimilarities are applied as distance metrics in two well known clustering algorithms, k-means and agglomerative hierarchical clustering algorithms. In this analysis, we also use internal and external cluster validity measures, aiming to compare the effectiveness of all four distance measures in both clustering algorithms.

Recent researches on feature selection have been conducted in an attempt to find efficient methods for automatic selection of relevant features. The idea is to select a subset of attributes which are as representative as possible of the original data. Committees of classifiers, also known as ensemble systems, are composed of individual classifiers, organized in a parallel way and their output are combined in a combination method, which provides the final output of the system. In the context of these systems, feature selection methods can be used to provide different subsets of attributes for the individual classifiers, aiming to reduce redundancy among the attributes of a pattern and to increase the diversity in such systems. There are several methods to select features in ensembles systems and genetic algorithms (GA) is one of the most used methods. The main problem of using GA is the choice of the fitness function since the use of the ensemble accuracy means a complex and time consuming process and filter approaches may not reflect the real meaning of the solution. In this paper, we use feature selection via genetic algorithm to generate different subsets for the individual classifiers. In our proposal, we will used a hybrid and adaptive fitness function, in which we consider both approaches, filter and wrapper. In order to evaluate our proposal, experiments were conducted involving 10 different types of machine learning algorithms on 14 datasets. We will analyse the performance results of the proposed model compared with a genetic algorithm using a filter approach as well as the standard Bagging algorithm without feature selection.

In traditional classification problems (single-label), patterns are associated with a single label from the set of disjoint labels. When an example can simultaneously belong to more than one label, we call it a multi-label classification problem. In relation to the learning strategy, the majority of classification methods requires a large number of training instances to be able to generalize the mapping function, making predictions with high accuracy. However, it is usually difficult to find a number of instances labeled which is sufficient to induce an accurate classification model. This problem is enhanced in the multi-label context, since the number of possible combinations in the label attributes increases considerably. In order to smooth out this problem, the idea of semi-supervised learning has emerged. It combines labeled and unlabeled data during the training phase. Some semi-supervised methods have been proposed for single-label classification methods. However, very little effort has been done in the context of multi-label classification. This paper proposes three semi-supervised methods for the multi-label classification. In order to validate the feasibility of these methods, an empirical analysis will be conducted, aiming to evaluate the performance of such methods in different tasks and using different evaluation metrics.

Although ensemble systems have been proved to be efficient for pattern recognition tasks, its elaboration and design is not an easy task. Some aspects such as the choice of its individual classifiers and the use of feature selection methods are very difficult to define. In addition, these aspects can have a strong effect in the accuracy of these systems, leading, for instance, to cases where the produced ensembles have no performance improvement. In order to avoid this situation, there is a great deal of research to select individual classifiers or distribute attributes to the individual classifiers of ensemble systems. In most of these works, however, only one aspect is tackled (either member selection or feature selection). In this paper, we present an analysis of two well-known optimization techniques to choose the ensemble members and to select attributes for these individual classifiers. In order to do this analysis, we use accuracy as well as two recently proposed diversity measures as parameters, in a multi-objective optimization problem.

Nowadays, the representation of many real word problems needs to use some type of relational model. As a consequence, information used by a wide range of systems has been stored in multi relational tables. However, from a data mining point of view, it has been a problem, since most of the traditional data mining algorithms have not been originally proposed to handle this type of data without discarding relationship information. Aiming to ameliorate this problem, we propose a hierarchical approach for handling relational data. In this approach the relational data is converted into a hierarchical structure (the main table as the root and the relations as the nodes). This hierarchical way to represent relational data can be used either for classification or clustering purposes. In this paper, we will use it in clustering algorithms. In order to do so, we propose a hierarchical distance metric to compute the similarity between the tables. In the empirical analysis, we will apply the proposed approach in two well-known clustering algorithms (k-means and agglomerative hierarchical). Finally, this paper also compares the effectiveness of our approach with one existing relational approach.

Bagging algorithm has been proven to be effective when dealing with on different classification problems. However, the success of Bagging depends strongly on the diversity level reached by the individual classifiers of the ensemble models. Diversity in ensemble can be obtained when the individual classifiers are built using different circumstances, such as parameter settings, training datasets and learning algorithms. This paper presents a new approach which combines these three different ways to obtain high diversity in Bagging models, aiming, as a consequence, to obtain high levels of accuracy for the ensembles. In the proposed approach, in order to obtain the optimal configurations of features and classifiers in Bagging models, we have applied an evolutionary approach composed of two genetic algorithm instances. In order to validate the proposed approach, experiments involving 10 classification algorithms have been conducted, applying the resulting Bagging structures in 5 pattern classification datasets taken from the UCI repository. In addition, we analyze the performance of the resulting Bagging structures in terms of two recently proposed diversity measures, referred to as good and bad.

Committees of classifiers, also known as ensemble systems, are composed of individual classifiers, organized in a parallel way and their output are combined in a combination method, which provides the final output of the system. In the context of these systems, feature selection methods can be used to provide different subsets of attributes for the individual classifiers, aiming to reduce redundancy among the attributes of a pattern and to increase the diversity in such systems. Since the problem of feature selection can be reduced to a search problem and that an exhaustive search for the subsets of attributes can be considered NP-hard, heuristic search can be adopted for solving this problem. This paper aims to introduce two important optimization techniques (Ant-colony and particle swarm) as a method to select attributes in an ensemble system as well as to compare their performance with Genetic Algorithm, whose research is well established in this area. These three algorithms have in common the fact that they bio-inspired meta-heuristics, since their search rules aim to simulate some aspects of the behavior of living beings.

In the context of ensemble systems, feature selection methods can be used to provide different subsets of attributes for the individual classifiers, aiming to reduce redundancy among the attributes of a pattern and to increase the diversity in such systems. Among the several techniques that have been proposed in the literature, optimization methods have been used to find the optimal subset of attributes for an ensemble system. In this paper, an investigation of two optimization techniques, genetic algorithm and ant colony optimization, will be used to guide the distribution of the features among the classifiers. This analysis will be conducted in the context of heterogeneous ensembles and using different ensemble sizes.

In traditional classification problems (single-label), patterns are associated with a single label from the set of disjoint labels (classes). When an example can simultaneously belong to more than one label, this classification problem is known as multi-label classification problem. Multi-label classification methods have been increasingly used in modern application, such as music categorization, functional genomics and semantic annotation of images. This paper presents a comparative analysis of some existing multi-label classification methods applied to different domains. The main aim of this analysis is to evaluate the performance of such methods in different tasks and using different evaluation metrics.

In the design of Classifier Ensembles, diversity is considered as one of the main aspects to be taken into account, since there is no gain in combining identical classification methods. One way of increasing diversity is to use feature selection methods in order to select subsets of attributes for the individual classifiers. In this paper, it is investigated the use of a simple reinforcement-based method, called ReinSel, in ensemble systems. More specifically, it is aimed to evaluate the capability of this method to select the correct attributes of a dataset, avoiding unimportant and noisy attributes.

Multi-classifier systems, also known as ensembles or committees, have been widely used to solve several classification problems, because they usually provide better performance than the individual classifiers. However, in order to build robust ensembles, it is necessary that the individual classifiers are as accurate as diverse among themselves - this is known as the diversity/accuracy dilemma. In this sense, some works analyzing the ensemble performance in context of this dilemma have been proposed. However, the majority of them address the homogenous structures of ensemble, i.e., ensembles composed only of the same type of classifiers. Thus, motivated by this limitation, this paper will perform an empirical investigation on the diversity/accuracy dilemma for heterogeneous ensembles. In order to do this, genetic algorithms will be used to guide the building of the ensemble systems.

Classifier ensembles, also known as committees, are systems composed of a set of base classifiers (organized in a parallel way) and a combination module, which is responsible for providing the final output of the system. The main aim of using ensembles is to provide better performance than the individual classifiers. In order to build robust ensembles, it is often required that the base classifiers are as accurate as diverse among themselves-this is known as the diversity/accuracy dilemma. There are, in the literature, some works analyzing the ensemble performance in context of such a dilemma. However, the majority of them address the homogenous structures, i.e., ensembles composed only of the same type of classifiers. Motivated by such a limitation, this paper presents an empirical investigation on the diversity/accuracy dilemma for heterogeneous ensembles. In order to do so, multi-objective genetic algorithms will be used to guide the building of the ensemble systems.

Classifier ensembles are systems composed of a set of individual classifiers (organized in a parallel way) and a combination module, which is responsible for providing the final output of the system. In the design of these systems, diversity is considered as one of the main aspects to be taken into account, since there is no gain in combining identical classification methods. One way of increasing diversity is to provide different datasets (patterns and/or attributes) for the individual classifiers. In this context, it is envisaged to use, for instance, feature selection methods in order to select subsets of attributes for the individual classifiers. However, the majority of the papers using feature selection for ensembles address the homogenous structures of ensemble, i.e., ensembles composed only of the same type of classifiers. In this paper, two approaches of genetic algorithms (single and multi-objective) will be used to guide the distribution of the features among the classifiers in the context of heterogeneous ensembles.

ARTMAP-based models are neural networks which uses a match-based learning procedure. The main advantage of ARTMAP-based models over error-based models, such as Multi-layer Perceptron, is the learning time, which is considered as significantly fast. This feature is extremely important in complex systems that require the use of several neural models, such as ensembles or committees, since they produce strong and fast classifiers. Aiming to add an extra contribution to ARTMAP-based ensembles, this paper presents an analysis of accuracy and diversity in these systems. As a result of this analysis, it is intended to detect any relation between these two parameters and to use this in the design of these systems.

Diversity is considered as one of the main prerequisites for an efficient use of ensemble systems. One way of increasing diversity is through the use of feature selection methods in ensemble systems. In this paper, a class-based feature selection method for ensemble systems is proposed. The proposed method is inserted into the filter approach of feature selection methods and it chooses only the attributes that are important only for a specific class. An analysis of the performance of the proposed method is also investigated in this paper and it shows that the proposed method has outperformed the standard feature selection method.

In systems that combine the outputs of classification methods (combination systems), such as ensembles and multi-agent systems, one of the main constraints is that the base components (classifiers or agents) should be diverse among themselves. In other words, there is clearly no accuracy gain in a system that is composed of a set of identical base components. One way of increasing diversity is through the use of feature selection or data distribution methods in combination systems. In this paper, an investigation of the impact of using data distribution methods among the components of combination systems will be performed. In this investigation, five different methods of data distribution will be used and an analysis of the combination systems, using several different configurations, will be performed. As a result of this analysis, it is aimed to detect which combination systems are more suitable to use feature distribution among the components.

This paper presents an investigation of the influence of the RePART (Reward and Punishment ARTmap) neural network in structures of ensembles designed by three variants of boosting: Aggressive, Conservative and Inverse Boosting. In this investigation, it is aimed to analyze whether the use of this model is positive for ARTMAP-based ensembles. In addition, it aims to define which boosting strategy is the most suitable to be used in ARTMAP-based ensembles.

In the context of Ensembles or Multi-Classifier Systems, the choice of the ensemble members is a very complex task, in which, in some cases, it can lead to ensembles with no performance improvement. In order to avoid this situation, there is a great deal of research to find effective classifier member selection methods. In this paper, we propose a selection criterion based on both the accuracy and diversity of the classifiers in the initial pool. Also, instead of using a static selection method, we use a Dynamic Classifier Selection (DSC) procedure. In this case, the member classifiers to form the ensemble are chosen at the test (use) phase. That is, different testing patterns can be classified by different ensemble configurations.

The determination of the three-dimensional (3D) structure of biological macromolecules at different configurations can be very important for understanding biological processes at the molecular level. The detection of individual particles from electron microscopy (EM) micrographs turns into a major labor-intensive bottleneck, when the number of particles needed starts to exceed a few tens of thousand molecular images. Multi-classifier systems have been widely investigated as tools for performing complex classifying tasks. In this work, we investigate the adequacy of using multi-classifier systems to detect particles on electron microscopy micrographs. In order to do so, we compare the performance of five algorithms for generating individual classifiers and three other ones for multi-classifier algorithms. Such results are also compared with others found in the literature. In terms of results, the multi-classifier systems generated show larger accuracy (correct classification) and lower false positive and negative rates.

Classifier Combination has been investigated as an alternative to obtain improvements in design and/or accuracy for difficult pattern recognition problems. In the literature, many combination methods and algorithms have been developed, including methods based on fuzzy sets. This paper presents an evaluation of how the choice of the components (classifiers) can affect the accuracy of some fuzzy combination methods. The main aim of this analysis is to investigate whether or not fuzzy combination methods are less affected by the choice of the ensemble members than the non-fuzzy methods.

The idea of including intelligent agents in the structure of multi-classifier systems has emerged in order to overcome some drawbacks of these systems and, as a consequence, to improve the performance of such systems. As a result of this, the ClassAge system was proposed. This system has presented good results in some classification tasks. In this paper, an extension of this system is presented. Basically, an optimization technique is used to optimize the functioning of ClassAge. Also, each ClassAge agent is composed of a group of classifiers, instead of one classifier that was used in the original version.

Feature selection methods are applied in ensembles in order to find subsets of features for the classifiers of the ensemble. The use of these methods aims to reduce the redundancy of the features as well as to increase diversity of the classifiers of an ensemble. In this paper, a comparative analysis of six different feature selection methods is performed in ensembles using six different combination methods. The main aim of this paper is to investigate which combination methods are more affected by the use of feature selection methods.

In the context of ensembles or multi-classifier systems, many combination methods and algorithms have been developed in the literature, including methods based on fuzzy sets. This paper presents an evaluation of how the choice of the components (classifiers) can affect the accuracy of some fuzzy combination methods. The main aim of this analysis is to investigate whether or not fuzzy combination methods are less affected by the choice of the ensemble members than the non-fuzzy methods.

The ClassAge system is a multi-agent system for classification tasks. This system was proposed as an attempt to include the idea of intelligent agents in the structure of multi- classifier systems (MCSs). Also, it is aimed to overcome some drawbacks of MCSs and, as a consequence, to improve the performance of such systems. In this paper, an extension of ClassAge is presented. Basically, an optimization technique is used to optimize the functioning of ClassAge. Also, each ClassAge agent is composed of a group of classifiers, instead of one classifier that was used in the original version.

The use of intelligent agents in the structure of multi-classifier systems has been investigated in order to overcome some drawbacks of these systems and, as a consequence, to improve the performance of such systems. As a result of this, the NeurAge system was proposed. This system is composed by several neural agents which communicate (negotiate) a common result for the testing patterns. The NeurAge system has been successfully applied in some classification tasks. Basically, in these investigations, NeurAge has used multi-layer perceptrons (MLPs) as the neural network module of its agents. In this paper, it is presented an investigation of the use of the NeurAge System using other types of classifiers, mainly Fuzzy MLP. The main aim of this investigation is to analyze the benefits of using fuzzy neural networks in the performance of the NeurAge System.

The use of intelligent agents in the structure of multi-classifier systems has been investigated in order to overcome some drawbacks of these systems and, as a consequence, to improve the performance of such systems. As a result of this, the NeurAge system was proposed. This system has presented good results in some conventional (centralized) classification tasks. Nevertheless, in some classification tasks, relevant features can be distributed over a set of agent. These applications can be classified as distributed classification tasks. In this paper, an investigation of the performance of the NeurAge system in distributed classification tasks will be performed. In other words, it will be investigated the performance of NeurAge when the data are distributed over its agents.

Ensemble of classifiers is an effective way of improving performance of individual classifiers. However, the task of selecting the ensemble members is often a non-trivial one. For example, in some cases, a bad selection strategy could lead to ensembles with no performance improvement. Thus, many researchers have put a lot of effort in finding an effective method for selecting classifier for building ensembles. In this context, a dynamic classifier selection (DCS) method is proposed, which takes into account both the accuracy and the diversity of the classifiers.

This paper presents a comparative analysis of some negotiation methods used in a multi-neural agent system (NeurAge). This systems is composed of several neural classifiers (called neural agents) and its main aim is to overcome some drawbacks of multi-classifier systems and, as a consequence, to improve performance of such systems.

This paper presents a wide evaluation of performance and diversity in hybrid and non-hybrid structures of ensembles. In applying some diversity measures at the chosen ensemble, it is intended to analyse the effect of varying diversity in ensembles and how the variation of diversity can affect the performance of several combination methods (selection-based and combination-based methods). Finally, it is also intended to understand the reasons that some combination methods are more affected by variation in diversity.

Protein fold recognition is an important approach to structure discovery without relying on sequence similarity. In this context, computer-based tools, mainly the techniques from machine learning (ML), have become essential considering the large volume of data. We present an empirical comparison of individual machine learning techniques (k-nearest neighbor, naive Bayes, decision trees, support vector machines and neural networks) and ensemble approaches (bagging and boosting) to the task of protein structural class prediction.

This paper investigates the performance of some multi-classifier systems, focusing on the benefits that can be gained when integrating different types of classifiers (hybrid multi-classifier systems). An empirical evaluation shows that the integration of different types of classifiers can lead to an improvement in performance in some practical classification tasks.

The most normal forms for fuzzy logics are versions of conjunctive and disjunctive classical normal forms. Unfortunately, they do not preserve neither 1-tautologies nor 0-contradictions. This paper introduces a normal form that partially preserves 1-tautologies for any continuous t-norm - i.e. if a formula is a 1-tautology then their normal form is also a 1-tautology but the reciprocal does not always hold. For the class of t-norms without zero divisors it preserves 0-contradictions, i.e. a formula is 0-contradiction if and only if their normal form is also 0-contradiction. The paper shows that this normal form could be used to implement an automatic theorem provers for a class of residuum-based propositional fuzzy logics.

This work presents a comparative investigation of the performance of the RePART model with the ARTMAP-IC model, when applied to two tasks of pattern recognition. Such investigation has as its main objective to verify whether the RePART model presents higher recognition rate than the performance presented by the other model. Also, a statistical method is applied to the results of the neural networks in order to analyse the significance from a statistical point of view of the neural network performances.

This paper investigates the performance of multi-neural systems, focusing on the benefits that can be gained when integrating different types of neural experts (hybrid multi-neural system). An empirical evaluation shows that the integration of different types of neural networks leads to an improvement in performance in a practical classification task for a range of combination methods.

The set of fuzzy connectives can be seen as an important combination tool, such as in combining the antecedent sets of the rules, in multi-criteria decision making and in combining the outputs of neural classifiers in a multi-neural system. This papers investigates the performance of some fuzzy combination schemes applied to a multi hybrid neural system which is composed of neural and fuzzy neural networks. An empirical evaluation in a handwritten numeral recognition task is used to investigate the performance of the presented fuzzy methods with some existing combination methods.

This paper presents a mechanism to vary the vigilance parameter in the RePART fuzzy neural network. This mechanism helps to smooth out the problem of category proliferation which affects ARTMAP-based networks. Empirical experiments show that the use of variable vigilance improves the performance of the RePART model while, at the same time, requiring a less complex structure.

It is well known that substantial improvements can be obtained in difficult pattern recognition problems by combining or integrating the outputs of multiple neural classifiers. This paper analyses the performance of some combination schemes applied to a multi-hybrid neural system which is composed of neural and fuzzy neural networks. Essentially, the combination methods employ different ways to extract valuable information from the output of the experts through the use of confidence (weights) measures of the ensemble members to each class. An empirical evaluation in a handwritten numeral recognition task is used to investigate the performance of the presented methods in comparison with some existing combination methods.