Search

Background Mammographic breast density is an important predictor of breast cancer, but its measurement has limitations related to subjectivity of visual evaluation or to difficult access for automatic volumetric measurement methods. Herein, we describe the design and clinical validation of Aguida, a software program for automated quantification of breast density from flat mammography images. Materials and methods The software program was developed in MatLab. After image segmentation separating the background from the breast image, the operator positions a cursor defining a region of interest on the pectoralis major muscle from the mediolateral oblique view. Then, in the craniocaudal view, the threshold for separation of the dense tissue is based on the optical density of the pectoral muscle, and the proportion of dense tissue is calculated by the program. Mammograms obtained from 2 different occasions in 291 women were used for clinical evaluation. Results The intraclass correlation coefficient (ICC) between breast density measurements by the software and by a radiologist was 0.96, with a bias of only 0.67 percentage points and a 95% limit of agreement of 13.5 percentage points; the ICC was 0.94 in the interobserver reliability assessment by two radiologists with different experience; and the ICC was 0.98 in the intraobserver reliability assessment. The distribution among the density classes was close to the values obtained with the volumetric software. Conclusions Measurement of breast density with the Aguida program from flat mammography images showed high agreement with the visual determination by radiologists, and high inter- and intra-observer reliability.

Given the actual volume and speed of financial transactions, financial fraud detection systems are constantly evolving based on new computational intelligence algorithms. Therefore, transaction monitoring and analysis prevent monetary losses caused by fraudsters. Since the fraud detection process is a labor-intensive task for human auditors given the huge amount of daily transactions processed by financial services information systems. Credit card is the financial product most explored in the financial fraud detection literature, while mobile money service is becoming a popular option for payments, fraud detection for such financial product has not yet been deeply explored. Therefore, it is interesting to optimize the auditing process and test new quantitative techniques, such as deep learning, to support human auditors before double-checking a suspicious transaction. Thus, we propose an integration of adversarial autoencoders and machine learning methods to perform an objective classification among three transaction types: regular, local, and global anomaly. The integration consists of using the autoencoder's generated latent vectors as features for the supervised learning algorithms. The experiments considered different latent vector space forms concerning their dimensionality and the clusters generated by a prior Gaussian mixture. The results show that some classifiers may accept latent characteristics well, getting better or similar performance when using all the original characteristics.

Several neural network models have been proposed in the literature to predict the future behavior of financial time series. However, an intrinsic limitation arises from this particular prediction task with modeling via neural networks, since the prediction, when sampled in daily frequency, have 1-step-ahead delay with respect to real time series observations. In order to overcome such drawback, we present a deep increasing–decreasing-linear neural network (wherein each layer is composed of a set of increasing–decreasing-linear processing units) to predict the behavior of financial time series. In addition, we present a learning process to train the proposed model using a descending gradient-based approach. In order to assess the model’s prediction performance, we use twelve financial time series from relevant stock markets around the world. The obtained results show that the proposed model have competitiveness, in terms of predictive performance, and have better effectiveness when compared to recent models presented in the literature of time series prediction.

In many regression and time series forecasting problems, the input data is not fully available at the beginning of the training phase. Conventional machine learning methods for batch data are not able to handle this problem. The sequential version of ELM, called Online Sequential Extreme Learning Machine (OS-ELM), addresses this problem through the least squares recursive solution for updating the network output weights. However, the implementation of OS-ELM and its extensions suffer from the problem of multicollinearity and its side effect on the variance of the weight estimates. This paper introduces a new method of sequential learning for handling the effects of multicollinearity. The proposed method, called Kalman Learning Machine (KLM), uses the Kalman filter to sequentially update the output weights of a Single Layer Feedforward Network (SLFN) based on OS-ELM. An extension of the proposed method, called Extended Kalman Learning Machine (EKLM), is presented in order to address the problem of nonlinear data. The proposed method was compared with some of the most recent and effective methods for handling the effects of multicollinearity in sequential learning problems. The experiments performed showed that the proposed method performs better than most state-of-the-art methods considering both the prediction error and training time.

Digital Libraries have been developed nowadays as a way to dispose digital information through the Internet. This is particularly very useful when the information comes from historical documents. This research takes place in the PROHIST Project [Mello et al., 2008] which aims the creation of a digital library with methods to preserve and broadcast images of historical documents. In general, the access to original documents has to be done carefully as, because of its age, the paper is more susceptible to the wear and tear over time. In order to make the documents more easily accessible, digitization comes as the most efficient solution. In a digital media, as digital images, the documents can be visualized and copied. This also helps the preservation of the documents as they are digitized in high resolution and in true color format. It is common to use JPEG file format (Sayood, 1996) to store these images ensuring a good space storage/quality ratio. However, even in this format, to access an archive of thousands of high quality true color images is not an easy task even with the extended use of broad band Internet. The storage space of the images can be reduced with its conversion to black-and-white images. In this bi-level format and stored using GIF file format, the size of the file can be five times lower than the original true color JPEG image. Binarization or thresholding (Parker, 1997) is the process that converts an image into black-and-white: a threshold value is defined and the colors above that value are converted into white, while the colors below it are converted into black. This is a very simple process in digital image processing when one has a document with black ink written on a white paper. Historical documents, however, have several types of noises. The degradation yellows the sheet of paper and creates some noise that is perceptible to the digitizing process. Even more, in some cases, the ink has faded. This is particularly important when the document is written on both sides of the paper. In some cases, the ink of one side interferes in the other creating an effect called “ink bleeding”. Because of these problems, it is very difficult to find the best threshold value that separates the colors that belong to the paper from the colors that belong to the ink. An example of such a document is presented in Figure 1-left. In this paper, we present a new thresholding algorithm for color quantization based on genetic algorithms and image fidelity metrics. These metrics are used to define the convergence point of the genetic algorithm. The quantized image is then binarized based on

The financial time series forecasting is considered a rather difficult problem, due to the many complex features frequently present in such time series (irregularities, volatility, trends and noise). Several approaches have been studied for the development of predictive models able to predict time series, based on its past and present data. In the attempt to solve the time series prediction problem, a wide number of linear statistical models were proposed. Among them, the popular linear statistical approach based on Auto Regressive Integrated Moving Average (ARIMA) models [1] is one of the most common choices. However, since the ARIMA models are linear and most real world applications involve nonlinear problems, this can introduce an accuracy limitation of the generated forecasting models. In the attempt to overcome linear statistical models limitations, other nonlinear statistical approaches have been developed, such as the bilinear models [2], the threshold autoregressive models [3], the exponential autoregressive models [4], the general state dependent models [5], amongst others. The drawbacks of those nonlinear statistical models are the high mathematical complexity associated with them (resulting in many situations in similar performances to the linear models) and the need, most of the time, of a problem dependent specialist to validate the predictions generated by the model, limiting the development of an automatic forecast system [6]. Alternately, Artificial Neural Networks (ANNs) based approaches have been applied for nonlinear modeling of time series in the last two decades [7-14]. However, in order to define a solution to a given problem, ANNs require the setting up of a series of system parameters, some of them are not always easy to determine. The ANN topology, the number of processing units, the algorithm for ANN training (and its corresponding variables) are just some of the parameters that require definition. In addition to those, in the particular case of time series forecasting, another crucial element necessary to determine is the relevant time lags to represent the series [15]. In this context, evolutionary approaches for the definition of neural network parameters have produced interesting results [16{20]. Some of these works have focused on the evolution of the network weights whereas others aimed at evolving the network architecture.

SigNet is a state of the art model for feature representation used for handwritten signature verification (HSV). This representation is based on a Deep Convolutional Neural Network (DCNN) and contains 2048 dimensions. When transposed to a dissimilarity space generated by the dichotomy transformation (DT), related to the writer-independent (WI) approach, these features may include redundant information. This paper investigates the presence of overfitting when using Binary Particle Swarm Optimization (BPSO) to perform the feature selection in a wrapper mode. We proposed a method based on a global validation strategy with an external archive to control overfitting during the search for the most discriminant representation. Moreover, an investigation is also carried out to evaluate the use of the selected features in a transfer learning context. The analysis is carried out on a writer-independent approach on the CEDAR, MCYT and GPDS datasets. The experimental results showed the presence of overfitting when no validation is used during the optimization process and the improvement when the global validation strategy with an external archive is used. Also, the space generated after feature selection can be used in a transfer learning context., arXiv admin note: text overlap with arXiv:2004.03373

This paper investigates the presence of overfitting when using Binary Particle Swarm Optimization (BPSO) to perform the feature selection in a context of Handwritten Signature Verification (HSV). SigNet is a state of the art Deep CNN model for feature representation in the HSV context and contains 2048 dimensions. Some of these dimensions may include redundant information in the dissimilarity representation space generated by the dichotomy transformation (DT) used by the writer-independent (WI) approach. The analysis is carried out on the GPDS-960 dataset. Experiments demonstrate that the proposed method is able to control overfitting during the search for the most discriminant representation.

In this paper, we argue that the unsatisfactory out-of-distribution (OOD) detection performance of neural networks is mainly due to the SoftMax loss anisotropy and propensity to produce low entropy probability distributions in disagreement with the principle of maximum entropy. Current out-of-distribution (OOD) detection approaches usually do not directly fix the SoftMax loss drawbacks, but rather build techniques to circumvent it. Unfortunately, those methods usually produce undesired side effects (e.g., classification accuracy drop, additional hyperparameters, slower inferences, and collecting extra data). In the opposite direction, we propose replacing SoftMax loss with a novel loss function that does not suffer from the mentioned weaknesses. The proposed IsoMax loss is isotropic (exclusively distance-based) and provides high entropy posterior probability distributions. Replacing the SoftMax loss by IsoMax loss requires no model or training changes. Additionally, the models trained with IsoMax loss produce as fast and energy-efficient inferences as those trained using SoftMax loss. Moreover, no classification accuracy drop is observed. The proposed method does not rely on outlier/background data, hyperparameter tuning, temperature calibration, feature extraction, metric learning, adversarial training, ensemble procedures, or generative models. Our experiments showed that IsoMax loss works as a seamless SoftMax loss drop-in replacement that significantly improves neural networks' OOD detection performance. Hence, it may be used as a baseline OOD detection approach to be combined with current or future OOD detection techniques to achieve even higher results., Accepted for publication in the IEEE Transactions on Neural Networks and Learning Systems: Special Issue on Deep Learning for Anomaly Detection

High number of writers, small number of training samples per writer with high intra-class variability and heavily imbalanced class distributions are among the challenges and difficulties of the offline Handwritten Signature Verification (HSV) problem. A good alternative to tackle these issues is to use a writer-independent (WI) framework. In WI systems, a single model is trained to perform signature verification for all writers from a dissimilarity space generated by the dichotomy transformation. Among the advantages of this framework is its scalability to deal with some of these challenges and its ease in managing new writers, and hence of being used in a transfer learning context. In this work, we present a white-box analysis of this approach highlighting how it handles the challenges, the dynamic selection of references through fusion function, and its application for transfer learning. All the analyses are carried out at the instance level using the instance hardness (IH) measure. The experimental results show that, using the IH analysis, we were able to characterize “good” and “bad” quality skilled forgeries as well as the frontier region between positive and negative samples. This enables futures investigations on methods for improving discrimination between genuine signatures and skilled forgeries by considering these characterizations.

The detection of objects considering a 6DoF pose is a common requirement to build virtual and augmented reality applications. It is usually a complex task which requires real-time processing and high precision results for adequate user experience. Recently, different deep learning techniques have been proposed to detect objects in 6DoF in RGB images. However, they rely on high complexity networks, requiring a computational power that prevents them from working on mobile devices. In this paper, we propose an approach to reduce the complexity of 6DoF detection networks while maintaining accuracy. We used Knowledge Distillation to teach portables Convolutional Neural Networks (CNN) to learn from a real-time 6DoF detection CNN. The proposed method allows real-time applications using only RGB images while decreasing the hardware requirements. We used the LINEMOD dataset to evaluate the proposed method, and the experimental results show that the proposed method reduces the memory requirement by almost 99\% in comparison to the original architecture with the cost of reducing half the accuracy in one of the metrics. Code is available at https://github.com/heitorcfelix/singleshot6Dpose., This paper was accepted by 2020 International Joint Conference on Neural Networks (IJCNN)

Most of the automatic fire alarm systems detect the fire presence through sensors like thermal, smoke, or flame. One of the new approaches to the problem is the use of images to perform the detection. The image approach is promising since it does not need specific sensors and can be easily embedded in different devices. However, besides the high performance, the computational cost of the used deep learning methods is a challenge to their deployment in portable devices. In this work, we propose a new deep learning architecture that requires fewer floating-point operations (flops) for fire recognition. Additionally, we propose a portable approach for fire recognition and the use of modern techniques such as inverted residual block, convolutions like depth-wise, and octave, to reduce the model's computational cost. The experiments show that our model keeps high accuracy while substantially reducing the number of parameters and flops. One of our models presents 71\% fewer parameters than FireNet, while still presenting competitive accuracy and AUROC performance. The proposed methods are evaluated on FireNet and FiSmo datasets. The obtained results are promising for the implementation of the model in a mobile device, considering the reduced number of flops and parameters acquired., Comment: Accepted in the IEEE International Joint Conference on Neural Networks (IJCNN), 2020

Data science concerns to look for ways for extracting knowledge from extraordinary volume of data in any domain. This paper introduces a new three-stage approach for time series forecasting based on quartile data that combines aggregation, prototype selection and forecasting in the Symbolic Data Analysis (SDA) framework. Initially, time series are summarized and classes of entities are obtained by representing new time series. A prototype selection process based on mutual information is applied in order to obtain a more representative class data set. Each class is described by a list of three continues values, called quartile symbolic data. Boxplots time series are displayed, as a special case. In this framework, symbolic data are obtained to take into account variability between the entities of the classes. Approaches for boxplots forecasting are constructed from multivariate statistical modeling. Synthetic and real symbolic time series are considered in the experiments to demonstrate the usefulness of the proposed approach.

A recent study has presented a model, called the increasing-decreasing-linear (IDL) model, which is able to efficiently predict the high-frequency stock market. Nevertheless, a drawback arises from the IDL's learning process, which consists of its costly methodology to circumvent the non-differentiability problem of increasing and decreasing operators. In this sense, trying to reduce the computational cost of the IDL design, we propose evolutionary learning machines, using the genetic algorithm, the particle swarm optimizer, the backtracking search algorithm, the firefly algorithm and the cuckoo search, to design the IDL model. Five relevant high-frequency time series from the Brazilian stock market are used to assess performance, and the achieved results have demonstrated better prediction performance with smaller computational cost when compared to those achieved by the IDL model designed by its classical learning process, as well as to those achieved by some relevant prediction models presented in the literature.

Supervised machine learning and opinion lexicon are the most frequent approaches for opinion mining, but they require considerable effort to prepare the training data and to build the opinion lexicon, respectively. In this paper, a novel unsupervised clustering approach is proposed for opinion mining. Three swarm algorithms based on Particle Swarm Optimization are evaluated using three corpora with different levels of complexity with respect to size, number of opinions, domains, languages, and class balancing. K-means and Agglomerative clustering algorithms, as well as, the Artificial Bee Colony and Cuckoo Search swarm-based algorithms were selected for comparison. The proposed swarm-based algorithms achieved better accuracy using the word bigram feature model as the pre-processing technique, the Global Silhouette as optimization function, and on datasets with two classes: positive and negative. Although the swarm-based algorithms obtained lower result for datasets with three classes, they are still competitive considering that neither labeled data, nor opinion lexicons are required for the opinion clustering approach.

Documents written in natural language constitute a major part of the artefacts produced during the software engineering life cycle. Studies indicate that more than 80% of enterprise data is stored in some sort of unstructured form, mainly as text. Therefore, the growth of user-generated content, especially from social media, provides a huge amount of data which allows discovering the experiences, opinions, and feelings of users. Text mining refers to the set of tools, techniques, and algorithms adopted to extract useful information from unstructured data. Considering that Portuguese ranks among the ten most spoken languages, and it is the second most common in Twitter, this study aims to map current primary studies that relate to the application of text mining for Portuguese. A systematic mapping method was applied and 6075 primary studies were retrieved up to the year 2014. A total of 203 studies were included, from which more than 60% analyse texts written in Brazilian variant. The majority of studies focus on the text classification task. Support vector machine and Naive Bayes appear as main the algorithms. Folha de Sao Paulo and Publico newspapers appear as main corpora, followed by the Portuguese Attorney General's Office corpus and Twitter.

The dilation-erosion-linear perceptron is a hybrid morphological neuron which has been recently proposed in the literature to solve some prediction problems. However, a drawback arises from such model for building mappings to solve tasks with complex input-output nonlinear relationships within effort estimation problems. In this sense, to overcome this limitation, we present a particular class of hybrid multilayer perceptrons, called the multilayer dilation-erosion-linear perceptron (MDELP), to deal with software development effort estimation problems. Each processing unit of the proposed model is composed of a mix between a hybrid morphological operator (given by a balanced combination between dilation and erosion operators) and a linear operator. According to Pessoa and Maragos’s ideas, we propose a descending gradient-based learning process to train the proposed model. Besides, we conduct an experimental analysis using relevant datasets of software development effort estimation and the achieved results are discussed and compared, according to MMRE and PRED25 measures, to those obtained by classical and state of the art models presented in the literature.

Identifying the language of a text is an important step for several natural language processing applications. State-of-the-art language identification (LID) systems perform very well when discriminating between unrelated languages on standard datasets. However, the LID task has a bottleneck when discriminating between similar languages or language varieties. Furthermore, LID has also proven to be very challenging when dealing with short texts such as the ones from Twitter. In this paper, we propose the use of smoothed n-gram language models to classify tweets in both Brazilian and European Portuguese variants. Word and character n-gram language models were combined and evaluated through five different classifiers. We have compared the smoothed n-gram language models together with the Term Frequency and Inverse Document Frequency weighting scheme. This paper also proposes an ensemble model, in which the class labels output were combined using majority voting and algebraic combiners. The best configuration reached accuracy of 92.71% using an ensemble model, which combines Lidstone (0.1) character 6-gram, Good–Turing word unigram, and Witten–Bell word bigram models, together with the Log-Likelihood Ratio estimation method.

Financial time series represent the stock prices over time and exhibit behavior similar to a data stream. Many works report on the use of data mining techniques to predict the future direction of stock prices and to discover patterns in the time series data to provide decision support for trading operations. Traditional optimization methods do not take into account the possibility that the function to be optimized, namely, the final financial balance for operations considering some stock, may have multiple peaks, i.e., be represented by multimodal functions. However, multimodality is a known feature of real-world financial time series optimization problems. To deal with this issue, this article proposes the PAA-MS-IDPSO-V approach (Piecewise Aggregate Approximation - Multi-Swarm of Improved Self-adaptive Particle Swarm Optimization with Validation). The proposed method aims to find patterns in financial time series to support investment decisions. The approach uses multi-swarms to obtain a better particle initialization for the final optimization phase since it aims to tackle multimodal problems. Furthermore, it uses a validation set with early stopping to avoid overfitting. The patterns discovered by the method are used together with investment rules to support decisions and thus help investors to maximize the profit in their operations in the stock market. The experiments reported in this paper compare the results obtained by the proposed model with the Buy-and-Hold, PAA-IDPSO approaches and another approach found in the literature. We report on experiments conducted with S&P100 index stocks and using the Friedman Non-Parametric Test with the Nemenyi post-hoc Test both with 95% confidence level. The results show that the proposed model outperformed the competing methods and was able to considerably reduce the variance for all stocks.

In this paper we present a study about the time series generator phenomenon from air pollutant concentration forecasting problems. It provides several evidences suggesting that this kind of generator phenomenon has nonlinear characteristics with long-term dependencies. Based on that, we present a nonlinear morphological model able to forecast these time series. Also, a descending gradient-based method, using ideas from the back-propagation algorithm, is presented to design the proposed model. Furthermore, an empirical analysis is conducted with the proposed model using a set of six relevant air pollutant concentration forecasting problems and employing three statistical measures to assess forecasting performance.

The dilation–erosion perceptron (DEP) is a hybrid morphological processing unit, composed of a balanced combination between dilation and erosion morphological operators, recently presented in the literature to solve some problems. However, a drawback arises from such model for building complex decision surfaces for non-linearly separable data. In this sense, to overcome this drawback, we present a particular class of morphological neural networks with multilayer structure, called the dilation–erosion neural network (DENN), to deal with binary classification problems. Each processing unit of the DENN is composed by a DEP processing unit. Also, a descending gradient-based learning process is presented to train the DENN, according to ideas from Pessoa and Maragos. Furthermore, we conduct an experimental analysis with the DENN using a relevant set of binary classification problems, and the obtained results indicate similar or superior classification performance to those achieved by classical and state of the art models presented in the literature.

The lack of labeled data and the need to update the learning model are the two main challenges when performing Opinion Mining with data streams. Usual Opinion Mining approaches are not able to deal with these challenges, nor with the characteristics brought by this kind of data. Moreover, the occurrence of changes (drifts) in the concepts and/or opinions is another issue. Possible alternative solutions to these problems are: the use of Semi-supervised learning, such as Active Learning, which labels selected data rather than the entire data set; or Multiple Classifiers Systems, which combines different classifiers and are well suited to deal with drifts. In this study, we combined these two approaches, proposing the use of eight Active Learning sampling strategies as a generation method for Multiple Classifiers Systems. The Active Learning approach requires the choice of a strategy to select the instances, and each strategy results in the creation of a distinct classification model. Our method was evaluated in 14 Twitter stream data sets and the results showed that it can be better for Opinion Mining with data streams than popular ensemble generation methods present in the literature, such as Bagging and AdaBoost.

Mobile devices are increasingly present in people's daily lives. However, despite the substantial improvement of new generations of smartphones, the amount of information and the complexity of the procedures delegated to these devices still imposes certain restrictions on processing, especially regarding energy consumption. A promise solution to this issue is the technique known as offloading. Over the last few years, several offloading support platforms have been proposed. This work has a particular interest in one of those platforms, called CAOS. Despite its success to perform offloading tasks, CAOS still has problems such as low scalability. In this paper, we report on the refactoring of CAOS into a new microservice-based architecture. Performance and scalability evaluations were performed in both monolithic and microservices versions to show the benefits achieved with the new CAOS architecture.

Out-of-distribution (OOD) detection approaches usually present special requirements (e.g., hyperparameter validation, collection of outlier data) and produce side effects (e.g., classification accuracy drop, slower energy-inefficient inferences). We argue that these issues are a consequence of the SoftMax loss anisotropy and disagreement with the maximum entropy principle. Thus, we propose the IsoMax loss and the entropic score. The seamless drop-in replacement of the SoftMax loss by IsoMax loss requires neither additional data collection nor hyperparameter validation. The trained models do not exhibit classification accuracy drop and produce fast energy-efficient inferences. Moreover, our experiments show that training neural networks with IsoMax loss significantly improves their OOD detection performance. The IsoMax loss exhibits state-of-the-art performance under the mentioned conditions (fast energy-efficient inference, no classification accuracy drop, no collection of outlier data, and no hyperparameter validation), which we call the seamless OOD detection task. In future work, current OOD detection methods may replace the SoftMax loss with the IsoMax loss to improve their performance on the commonly studied non-seamless OOD detection problem., Accepted for publication in The International Joint Conference on Neural Networks (IJCNN), 2021

There are two main problems when performing Opinion Mining (OM) with data streams: the lack of labeled data and the need to update the learning model. The most used OM techniques cannot deal well with these challenges, so, an alternative is to use semi-supervised methods, such as the Active Learning, which is a method to label only selected data rather than the entire data set; however, it requires the choice of a sampling strategy to select the data to be labeled. In this paper, we evaluated eight strategies in ten data sets, in order to identify the best ones for OM with Twitter streams. According to our experiments, the Entropy strategy showed the best results, but it selects a large number of instances to be labeled, requiring further investigation.

Concept drift is a change in the joint probability distribution of the problem. This term can be subdivided into two types: real drifts that affect the conditional probabilities p(y|x) or virtual drifts that affect the unconditional probability distribution p(x). Most existing work focuses on dealing with real concept drifts. However, virtual drifts can also cause degradation in predictive performance, requiring mechanisms to be tackled. Moreover, as virtual drifts frequently mean that part of the old knowledge remains useful, they require different strategies from real drifts to be effectively tackled. Motivated on this, we propose an approach called Gaussian Mixture Model for Dealing With Virtual and Real Concept Drifts (GMM-VRD), which updates and creates Gaussians to tackle virtual drifts and resets the system to deal with real drifts. The main results show that the proposed approach obtained the best results, in terms of average accuracy, in relation to the literature methods, which propose to solve that same problem. In terms of accuracy over time, the proposed approach showed lower degradation on concept drifts, which indicates that the proposed approach was efficient.

When compared to Writer-Dependent (WD) Handwritten Signature Verification, in which a model is trained for each individual writer, the Writer-Independent (WI) approach offers greater scalability, since only a single model is trained for all users from a dissimilarity space generated by the dichotomy transformation. However, many samples from the dissimilarity space are redundant and have little influence during the training of the verification model. This work investigates whether prototype selection (PS) preprocessing can be used in the space resulting from the dichotomy transformation without degrading the performance of the classifier. Furthermore, an investigation is also performed to examine the use of a WI classifier in a transfer learning scenario, i.e., where the classifier is trained in one dataset, and is used to verify signatures in other datasets. The experiments reported herein show that the use of prototype selection in the dissimilarity space allows a reduction in the complexity of the classifier without degrading its generalization performance. In addition, the results show that the WI classifier is scalable enough to be used in a transfer learning approach, with a resulting performance comparable to that of a classifier trained and tested in the same dataset. An analysis of the results obtained based on the instance hardness (IH) measure and dendrogram diagrams is performed in order to better understand the behavior of the resulting dichotomy transformation.

Politics is a commonly used domain in Opinion Mining applications, in which opinions may change over time. Nevertheless, the usual approaches for Opinion Mining are not able to deal with the characteristics and the challenges brought by continuous data streams; so, an alternative is the use of techniques such as Active Learning, which labels selected data rather than the entire data set. The Active Learning approach requires the choice of a sampling strategy to select the most valuable instances. However, no study has performed an analysis in order to identify the best strategies for Opinion Mining. In this sense, we evaluated eight Active Learning sampling strategies, from which Entropy achieved the best results. In addition, due to the lack of publicly available stream data sets written in Portuguese, we created and evaluated corpora from Twitter and Facebook about the 2018 Brazilian presidential elections.

The Software Effort Estimation (SEE) process has been approached in different ways in the literature, including models built from Machine Learning (ML). The combination of these models (Ensemble) is an important research topic in ML, and has lead to improvements in accuracy compared to individual models. This paper proposes heterogeneous and dynamic ensemble selection (DES) models, composed by a set of regressors dynamically selected by classifiers to estimate software development effort. In the training phase, a pool of regression algorithms is trained using training data and a validation data set to validate the models. Next, some classifiers are trained to identify the best regression model from the pool for each training instance. In the test phase each trained classifier is used to dynamically select a regressor model from the pool for predicting the effort for each test instance. The final prediction is given by the combination of the predictions of the regressors selected by the classifiers. An experimental analysis considering a relevant set of software effort estimation problems is reported. The experiments demonstrate that the proposed method outperforms individual regressors and some state of the art models of the literature.

Real-world applications have been dealing with large amounts of data that arrive over time and generally present changes in their underlying joint probability distribution, i.e., concept drift. Concept drift can be subdivided into two types: virtual drift, which affects the unconditional probability distribution p(x), and real drift, which affects the conditional probability distribution p(y|x). Existing works focuses on real drift. However, strategies to cope with real drift may not be the best suited for dealing with virtual drift, since the real class boundaries remain unchanged. We provide the first in depth analysis of the differences between the impact of virtual and real drifts on classifiers' suitability. We propose an approach to handle both drifts called On-line Gaussian Mixture Model With Noise Filter For Handling Virtual and Real Concept Drifts (OGMMF-VRD). Experiments with seven synthetics and seven real-world datasets show that OGMMF-VRD outperforms other approaches with separate mechanisms to deal with virtual and real drifts. It also has more stable rankings and smaller drops in performance during drifting periods than existing ensemble approaches, thus being more reliable for adoption in practice.

We propose a survey of soft computing techniques applied to financial market.We surveyed several primary studies proposed in the literature.A framework for building an intelligent trading system was proposed.Future directions of this research field are discussed. Financial markets play an important role on the economical and social organization of modern society. In these kinds of markets, information is an invaluable asset. However, with the modernization of the financial transactions and the information systems, the large amount of information available for a trader can make prohibitive the analysis of a financial asset. In the last decades, many researchers have attempted to develop computational intelligent methods and algorithms to support the decision-making in different financial market segments. In the literature, there is a huge number of scientific papers that investigate the use of computational intelligence techniques to solve financial market problems. However, only few studies have focused on review the literature of this topic. Most of the existing review articles have a limited scope, either by focusing on a specific financial market application or by focusing on a family of machine learning algorithms. This paper presents a review of the application of several computational intelligent methods in several financial applications. This paper gives an overview of the most important primary studies published from 2009 to 2015, which cover techniques for preprocessing and clustering of financial data, for forecasting future market movements, for mining financial text information, among others. The main contributions of this paper are: (i) a comprehensive review of the literature of this field, (ii) the definition of a systematic procedure for guiding the task of building an intelligent trading system and (iii) a discussion about the main challenges and open problems in this scientific field.

Efficiency in the operation is a crucial element for organizations, and this requires greater knowledge about existing business processes. Process Mining is an emerging discipline which aims to provide knowledge about business processes through learning from event logs from the information systems of the organization. Over the last years, the field has progressed significantly, however, some important challenges remain. Among these, we highlight the application in real-world environments, because organizations do not always have well-structured processes, that is, repeatable activities whose inputs and outputs are well defined. Also, there are more complex processes that make it more difficult to get knowledge. This paper presents a study, performed in a real environment, to evaluate the challenges and limitations of the Process Mining tools on process discovery.

The simplest linear combination of time series forecasters has shown a better performance than individual models. Thus, many researchers have sought to combine models for improving the forecasting process. This paper introduces a copulas-based approach (CB) for aggregating forecasters. The CB is based on the Cacoullos copula, with architecture divided into three parts: Single Modelling, Marginal Probability Distribution Computational and Joint Probability Distribution Computation. In the first part, the forecasts of individual models are obtained. In the second part, the residuals of the individuals models are calculated. In the third part, the models are combined via Cacoullos copula, based on the obtained residuals. The paper also evaluates the performance of the CB via simulated as well as financial time series (e.g. Google Stock Value). Thus, a comparative analysis is presented between CB and individual models (e.g. Artificial Neural Networks) and alternative combined forecasters (e.g. Simple Average-SA and Normal copula-CN). This study showed that the CB model produces better results when compared with the individual models, SA and CN.

Semantic segmentation is an essential task in computer vision that aims to label each image pixel. Several of the actual best approaches in this context are based on deep neural networks. For example, SegNet is a deep encoderdecoder architecture approach whose results were disruptive because it is fast and performs well. However, this architecture fails to finedelineating the edges between the objects of interest in the image. We propose some modifications in the SegNet-Basic architecture by using a post-processing segmentation layer (using Conditional Random Fields) and by transferring high resolution features combined to the decoder network. The proposed method was evaluated in the dataset CamVid. Moreover, it was compared with important variants of SegNet and showed to be able to improve the overall accuracy of SegNet-Basic by up to 17.5%.

Current research in the field of convolutional neu-ral networks usually focuses on improving network accuracy, regardless of the network size and inference time. In this paper, we investigate the effects of storage space reduction in SqueezeNet as it relates to inference time when processing single test samples. In order to reduce the storage space, we suggest adjusting SqueezeNet’s Fire Modules to include Depthwise Separable Convolutions (DSC). The resulting network, referred to as SqueezeNet-DSC, is compared to different convolutional neural networks such as MobileNet, AlexNet, VGG19, and the original SqueezeNet itself. When analyzing the models, we consider accuracy, the number of parameters, parameter storage size and processing time of a single test sample on CIFAR−10 and CIFAR-100 databases. The SqueezeNet-DSC exhibited a considerable size reduction (37% the size of SqueezeNet), while experiencing a loss in network accuracy of 1,07% in CIFAR−10 and 3,06% in top 1 CIFAR−100.

The use of features extracted using a deep convolutional neural network (CNN) combined with a writer-dependent (WD) SVM classifier resulted in significant improvement in performance of handwritten signature verification (HSV) when compared to the previous state-of-the-art methods. In this work it is investigated whether the use of these CNN features provide good results in a writer-independent (WI) HSV context, based on the dichotomy transformation combined with the use of an SVM writer-independent classifier. The experiments performed in the Brazilian and GPDS datasets show that (i) the proposed approach outperformed other WI-HSV methods from the literature, (ii) in the global threshold scenario, the proposed approach was able to outperform the writer-dependent method with CNN features in the Brazilian dataset, (iii) in an user threshold scenario, the results are similar to those obtained by the writer-dependent method with CNN features.

In this paper, a new sequential learning algorithm is constructed by combining the Online Sequential Extreme Learning Machine (OS-ELM) and Kalman filter regression. The Kalman Online Sequential Extreme Learning Machine (KOSELM) handles the problem of multicollinearity of the OS-ELM, which can generate poor predictions and unstable models. The KOSELM learns the training data one-by-one or chunk-by-chunk by adjusting the variance of the output weights through the Kalman filter. The performance of the proposed algorithm has been validated on benchmark regression datasets, and the results show that KOSELM can achieve a higher learning accuracy than OS-ELM and its related extensions. A statistical validation for the differences of the accuracy for all algorithms is performed, and the results confirm that KOSELM has better stability than ReOS-ELM, TOSELM and LS-IELM.

A model to overcome the random walk dilemma for high-frequency financial time series.A gradient-based training algorithm with automatic time phase adjustment.An experimental analysis using time series from Brazilian high-frequency stock market. Several models have been presented to solve the financial time series forecasting problem. However, even with sophisticated techniques, a dilemma arises from all these models, called the random walk dilemma (RWD). In this context, the concept of time phase adjustment was proposed to overcome such dilemma for daily frequency financial time series. However, the evolution of trading platforms had increased the frequency for performing operations on the stock market to fractions of seconds, which makes needed the analysis of high-frequency financial time series. In this way, this work presents a model, called the increasing decreasing linear neuron (IDLN), for high-frequency stock market forecasting. Also, a descending gradient-based method with automatic time phase adjustment is presented for the proposed model design. Besides, an experimental analysis is conduced using a set of high-frequency financial time series from the Brazilian stock market, and the achieved results overcame those obtained by establishing forecasting models in the literature.

Software development effort estimation is the process of predicting the effort required to develop a software system. In order to improve the estimation accuracy, many different models have been proposed in the literature. Multiple classification systems represent an important field of research for machine learning. In order to estimate software development effort, this paper proposes a heterogeneous and dynamic ensemble selection model, composed by a set of regressors dynamically selected by classifiers. Along with the proposed method it is conducted an experimental analysis involving a relevant set of software effort estimation problems, which has led to better results than those achieved by classical and state of the art models previously presented.

Time series forecasting is a problem with many applications. However, in many domains, such as stock market, the underlying generating process of the time series observations may change, making forecasting models obsolete. This problem is known as Concept Drift. Approaches for time series forecasting should be able to detect and react to concept drift in a timely manner, so that the forecasting model can be updated as soon as possible. Despite the fact that the concept drift problem is well investigated in the literature, little effort has been made to solve this problem for time series forecasting so far. This work proposes two novel methods for dealing with the time series forecasting problem in the presence of concept drift. The proposed methods benefit from the Particle Swarm Optimization (PSO) technique to detect and react to concept drifts in the time series data stream. It is expected that the use of collective intelligence of PSO makes the proposed method more robust to false positive drift detections while maintaining a low error rate on the forecasting task. Experiments show that the methods achieved competitive results in comparison to state-of-the-art methods.

In this work, we investigate speaker-specific filter banks for text-independent speaker verification. The proposed method performs an heuristic search for the best filter-bank configuration using the Artificial Bee Colony (ABC) algorithm and a proper fitness function for the standard i-vectors/PLDA-based speaker verification system. Furthermore, filter-bank decorrelated amplitudes are used instead of the cepstral coefficients produced by Discrete Cosine Transform (DCT). In the experiments, the proposed method is compared to standard Mel and linear scales in both cases where the decorrelation is performed using DCT and high-pass filtering. The comparison is performed on the MIT Mobile Device Speaker Verification Corpus in a gender-dependent trial scheme. The proposed method outperformed the baseline systems in almost all the test sets for both genders. Performance gains of 4.6% and 26.0% are achieved for male and female speakers, respectively.

This paper makes use of diverse domains’ datasets to analyze the impact of image complexity and diversity on the task of transfer learning in deep neural networks. As the availability of labels and quality instances for several domains are still scarce, it is imperative to use the knowledge acquired from similar problems to improve classifier performance by transferring the learned parameters. We performed a statistical analysis through several experiments in which the convolutional neural networks (LeNet-5, AlexNet, VGG-11 and VGG-16) were trained and transferred to different target tasks layer by layer. We show that when working with complex low-quality images and small datasets, fine-tuning the transferred features learned from a low complexity source dataset gives the best results.

Novelty detection is a problem with a large number of relevant applications. For some applications, the main interest is the prevention or detection of undesired states. In some cases, these undesired states are not known in advance; in others, such as machine monitoring, for instance, machine breakdown may be very rare and examples of this event may be unavailable. In such cases, the most widely accepted approach is to model the normal behavior of the system in order to subsequently detect unknown events. This is the basic concept of One-Class Classification (OCC). In some preliminary works we have proposed the Feature Boundaries Detector for One-Class Classification FBDOCC method, which operates by examining each problem feature at a time. In this paper we propose an extensive study of the behavior of the FBDOCC and introduce another version of the method, namely FBDOCC2. This work, also considers the use of the Particle Swarm Optimization (PSO) algorithm to find the best parameter configuration of the proposed method. Furthermore, this work also introduces a procedure to improve the training time performance without degrading the quality of the classification as well as some other contributions hereafter described. A number of experiments were carried out with synthetic and real datasets aiming at comparing both versions of the FBDOCC method with the most recent and effective OCC methods, namely: SVDD, One-class SVM, Least Squares One-class SVM, Kernel PCA, Gaussian Process Prior OCC, Condensed Nearest Neighbor Data Description and One-class Random Forests. The evaluation metrics considered in the experiments are: (i) the Area Under the ROC Curve (AUC); (ii) the Matthews Correlation Coefficient (MCC); (iii) the training time; and (iv) the prototype reduction rate. Regarding the metrics AUC and MCC, the first FBDOCC version has presented the best overall results among all the methods whereas the FBDOCC2 obtained results comparable to the best methods in some experiments where the standard FBDOCC yielded a poorer performance. FBDOCC was the faster method to train in comparison to the other methods in all but one dataset. In addition, FBDOCC was much faster than all methods based on support vector machines.

Supervised machine learning (ML) and lexicon-based are the most frequent approaches for opinion mining (OM), but they require considerable effort for preparing the training data and to build the opinion lexicon, respectively. This paper presents two unsupervised approaches for OM based on Particle Swarm Optimization (PSO). The PSO-based approaches were evaluated by eighteen experiments with different corpora types, domains, language, class balancing and pre-processing techniques. The proposed approaches achieved better accuracy on twelve experiments. Best results were obtained on corpora with a reduced number of dimensions and for specific domains. Best accuracy (0.79) was obtained by Discrete IDPSO on the OBCC corpus, outperforming supervised ML and lexicon-based approaches for this corpus.

Twitter is one of the most used social media with users generating about 1 million messages per day. As a result of the expansion of this microblog, there is a diversity of languages used by users and many studies aimed at identifying the language of tweets. The third most used language on Twitter is Portuguese, a pluricentric language with two national standard varieties: Brazilian Portuguese and European Portuguese. Identifying a language variety may positively impact various Natural Language Processing tasks, but accomplishing this task is still regarded as one of the bottlenecks in this area, especially when combined with another bottleneck, language identification applied to short texts. Thus, given these challenges, this paper provides a current view on the automatic discrimination of the two main Portuguese language varieties on Twitter texts by using an acknowledged approach with different techniques and features in order to get an optimum configuration to fit our problem. Results reached 0.9271 for accuracy using an ensemble method, which combines character 6-grams and word unigrams and bigrams.

One-class classification (OCC) is an important problem with applications in several different areas such as outlier detection and machine monitoring. Since in OCC there are no examples of the novelty class, the description generated may be a tight or a bulky description. Both cases are undesirable. In order to create a proper description, the presence of examples of the novelty class is very important. However, such examples may be rare or absent during the modeling phase. In these cases, the artificial generation of novelty samples may overcome this limitation. In this work it is proposed a two steps approach for generating artificial novelty examples in order to guide the parameter optimization process. The results show that the adopted approach has shown to be competitive with the results achieved when using real (genuine) novelty samples.

A time series is a sequence of observations collected over fixed sampling intervals. Several real-world dynamic processes can be modeled as a time series, such as stock price movements, exchange rates, temperatures, among others. As a special kind of data stream, a time series may present concept drift, which affects negatively time series analysis and forecasting. Explicit drift detection methods based on monitoring the time series features may provide a better understanding of how concepts evolve over time than methods based on monitoring the forecasting error of a base predictor. In this paper, we propose an online explicit drift detection method that identifies concept drifts in time series by monitoring time series features, called Feature Extraction for Explicit Concept Drift Detection (FEDD). Computational experiments showed that FEDD performed better than error-based approaches in several linear and nonlinear artificial time series with abrupt and gradual concept drifts.

Efforts have been made in financial markets to deal with price movement predicting. Recent studies have shown that the market can be outperformed by methodologies with the aid of science. In other words, it has been shown that methods based on computational intelligence can be more profitable than a buy-and-hold strategy. This paper proposes a probabilistic and dynamic chart pattern recognition hybrid system, by using soft computing techniques, such as Linearized Fuzzy C-Medoids (LFCMdd), Perceptually Important Points (PIP), Dynamic Time Warping (DTW) and Probabilistic Support Vector Machines (PSVM), to predict three important foreign exchange rates. Also, comparisons are performed against trading systems built with the use of technical indicators, optimized by Genetic Algorithms (GA). The experiments were performed with the sliding window strategy. The results showed that the proposed methodology outperformed the GA-based methodology, with mean return (ROI) of 277.6% against 20.96% for the GA-based.

The growth of social media and user-generated content (UGC) on the Internet provides a huge quantity of information that allows discovering the experiences, opinions, and feelings of users or customers. Opinion Mining (OM) is a sub-field of text mining in which the main task is to extract opinions from UGC. Given that Portuguese is one of the most common spoken languages in the world, and it is also the second most frequent on Twitter, the goal of this work is to plot the landscape of current studies that relates the application of OM for Portuguese. A systematic mapping review (SMR) method was applied to search, select and to extract data from the included studies. Manual and automated searches retrieved 6075 studies up to year 2014, from which 25 articles were included. Almost 70 % of all approaches focus on the Brazilian Portuguese variant. Naive Bayes and Support Vector Machine were the main classifiers and SentiLex-PT was the most used lexical resource. Portugal and Brazil are the main contributors in processing the Portuguese language.

Unstructured data accounts for more than 80 % of enterprise data and is growing at an annual exponential rate of 60 %. Text mining refers to the process of discovering new, previously unknown and potentially useful information from a variety of unstructured data including user-generated text content (UGTC). Given that Portuguese language is one of the most common languages in the world, and it is also the second most frequent language on Twitter, the goal of this work is to plot the landscape of current studies that relates the application of text mining to UGTC in the Portuguese language. The systematic mapping review method was applied to search, select, and to extract data from the included studies. Our manual and automated searches retrieved 6075 studies up to year 2014, from which 35 were included in the study. Text classification concentrates 79 % of all text mining tasks, having the Naive Bayes as the main classifier and Twitter as the main data source.