Search

This paper presents an approach to investigate distributed machine learning workloads on Spark. The work analyzes hardware and volume data factors regarding time and cost performance when applying machine learning (ML) techniques in big data scenarios. The method is based on the Design of Experiments (DoE) approach and applies randomized two-level fractional factorial design with replications to screening the most relevant factors. A Web Corpus was built from 16 million webpages from Portuguese-speaking countries. The application was a binary text classification to distinguish Brazillian Portuguese from other variations. Five different machine learning algorithms were examined: Logistic Regression, Random Forest, Support Vector Machines, Naive Bayes and Multilayer Perceptron. The data was processed using real clusters having up to 28 nodes, each composed of 12 or 32 cores, 1 or 7 SSD disks, and 3x or 6x RAM per core, totalizing a maximum computational power of 896 cores and 5.25 TB RAM. Linear models were applied to identify, analyze and rank the influence of factors. A total of 240 experiments were carefully organized to maximize the detection of non-cofounded effects up to the second-order, minimizing the experimental efforts. Our results include linear models to estimate time and cost performance, statistical inferences about effects, and a visualization tool based on parallel coordinates to aid decision making about cluster configuration.

Over-sampling algorithms are the most adopted approach to balance class distribution in imbalanced data problems, through random replication or synthesis of new examples in the minority class. Current over-sampling algorithms, however, usually use all available examples in the minority class to synthesise new instances, which may include noisy or outlier data. This work proposes k-INOS, a new algorithm to prevent over-sampling algorithms from being contaminated by noisy examples in the minority class. k-INOS is based on the concept of neighbourhood of influence and works as a wrapper around any over-sampling algorithm. A comprehensive experimentation was conducted to test k-INOS in 50 benchmark data sets, 8 over-sampling methods and 5 classifiers, with performance measured according to 7 metrics and Wilcoxon signed-ranks test. Results showed, particularly for weak classifiers (but not only), k-INOS significantly improved the performance of over-sampling algorithms in most performance metrics. Further investigations also allowed to identify conditions where k-INOS is likely to increase performance, according to features and rates measured from the data sets. The extensive experimentation framework evidenced k-INOS as an efficient algorithm to be applied prior to over-sampling methods.

The main contributions of this paper are: A new method for recognition of non-deformable severe occluded objects.Local shape descriptor from the wavelet coefficients of the tangent angle signature.Most likely object retrieval based on an ensemble of Hidden Markov Models.Two new area restrictions for consistency checking between hypotheses and occlusion. Shape classification has multiple applications. In real scenes, shapes may contain severe occlusions, hardening the identification of objects. In this paper, a method for object recognition under severe occlusion is proposed. Occlusion is dealt with separating shapes into parts through high curvature points, then tangent angle signature is found for each part and continuous wavelet transform is calculated for each signature. Next, the best matching object is calculated for each part using Pearsons correlation coefficient as similarity measure between wavelets of the part and of the most probable objects in the database. For each probable class, a hidden Markov model is created in an ensemble through training with the one-class approach. For hypotheses validation, two area restrictions are set to enhance recognition performance. Experiments were conducted with 1500 test shape instances with different scenarios of object occlusion. Results showed the method not only was capable of identifying highly occluded shapes (60%80% overlapping) but also present several advantages over previous methods.

A method based on hidden Markov models is described capable of dealing with severe part occlusions in different object recognition situations. Occlusion is dealt with separating shapes into parts through high curvature points, tangent angle signatures for each part and continuous wavelet transform for signatures. A hidden Markov model is created for each probable class in an ensemble trained with database objects from each class example. A strategy for search space sorting and stopping criterion are introduced to enhance retrieval performance of hypotheses from the database. The influence of sorting with proposed method is analysed to reduce recognition time on MPEG-7 and CMU KO datasets. Results showed the method not only was capable of identifying highly occluded shapes but also has relevant advantages over typical non-sorting approaches, reducing significantly the search for best hypotheses while keeping high recognition rates.

This paper proposes a method to perform time series prediction based on perturbation theory. The approach is based on continuously adjusting an initial forecasting model to asymptotically approximate a desired time series model. First, a predictive model generates an initial forecasting for a time series. Second, a residual time series is calculated as the difference between the original time series and the initial forecasting. If that residual series is not white noise, then it can be used to improve the accuracy of the initial model and a new predictive model is adjusted using residual series. The whole process is repeated until convergence or the residual series becomes white noise. The output of the method is then given by summing up the outputs of all trained predictive models in a perturbative sense. To test the method, an experimental investigation was conducted on six real world time series. A comparison was made with six other methods experimented and ten other results found in the literature. Results show that not only the performance of the initial model is significantly improved but also the proposed method outperforms the other results previously published.

This work describes an award winning approach for solving the NN3 Forecasting Competition problem, focusing on the sound experimental validation of its main innovative feature. The NN3 forecasting task consisted of predicting 18 future values of 111 short monthly time series. The main feature of the approach was the use of the median for combining the forecasts of an ensemble of 15 MLPs to predict each time series. Experimental comparison to a single MLP shows that the ensemble increases the performance accuracy for multiple-step ahead forecasting. This system performed well on the withheld data, having finished as the second best solution of the competition with an SMAPE of 16.17%.

The Time-delay Added Evolutionary Forecasting (TAEF) approach is a new method for time series prediction that performs an evolutionary search for the minimum number of dimensions necessary to represent the underlying information that generates the time series. The methodology proposed is inspired in Takens theorem and consists of an intelligent hybrid model composed of an artificial neural network combined with a modified genetic algorithm. Initially, the TAEF method finds the best fitted model to forecast the series and then performs a behavioral statistical test in order to adjust time phase distortions that may appear in the representation of some series. An experimental investigation conducted with relevant time series show the robustness of the method through a comparison, according to several performance measures, to previous results found in the literature and those obtained with more traditional methods.

Artificial neural networks (ANN) and related algorithms have been successfully applied to many pattern classification applications. In credit risk analysis, simple networks such as the multilayer perceptron (MLP) working alone and as ensembles have shown attractive performances when compared to more traditional statistical techniques and other market approaches. Despite that, new individual or combined models continue to be investigated to improve classifier performance in complex data problems involving large volumes of data, many correlated variables, and, noisy, incomplete, and inconsistent data. A particular attention for improving classifier accuracy should be given to the responses provided by the classifier in the frontiers that separate one class from the other. Once defining a border between two or more classes, it is expected that some of the cases very close to the border zones are incorrectly classified, due to the proximity of their pattern characteristics in that area. The purpose of this work is to investigate how a model based on case based reasoning (CBR) can be constructed to carry out a re-classification process for patterns around class frontiers. It is shown how the model performs in a large-scale credit risk assessment problem with data gathered from the PAKDD 2009 credit risk analysis competition. The performance of the CBR model combined with a MLP network is compared to that achieved by the MLP network alone with a 99% confidence level and it is shown that the model studied is promising when dealing with boundary data patterns.

The reliability of feedforward neural networks with respect to the rejection of patterns not belonging to the defined training classes is investigated. It is shown how networks with different activation functions and propagation rules construct the decision regions in the pattern space and, therefore, affect the network's performance in dealing with spurious information. A modification to the standard MLP structure is described to enhance its reliability in this respect.

In the past few years, artificial immune systems (AIS) have been proposed as an alternative approach to solve computational intelligence problems. Despite their interesting properties and theoretical appeal, the AIS algorithms still require a broader experimentation in real and large scale problems to ensure that they can approximate or outperform the more traditional techniques. In this work, the two known AIS models SAIS and CLONALG are studied in the problem of credit risk analysis making use of three benchmarking databases for comparative analysis. Further variations of the initially proposed algorithms are experimented and their performance are compared to classic logistic regression with respect to the KS (kolmogorov-Smirnov) test and ROC curves. Also, a method for score computation is introduced for allowing a more robust analysis through the ROC and KS metrics. The results can shed some light into the potential of using AIS models in daily real-world credit risk assessment operations.

This paper presents an approach for solving WCCI 2008's Ford Classification Challenge Problem. The solution is based on the creation of new input variables through temporal feature extraction and on the combination via bagging of an ensemble of 30 multi-layer perceptrons trained on sets divided by multiple random sampling of the labeled data. Signal power, signal to noise ratio and signal frequency were some of the meaningful features extracted for improving the system's performance. The data sampling strategy produced a robust median MLP response and allowed for the definition of the appropriate decision threshold. The performance measured on the 30 test samples (statistically independent from the training data) reached an average of Max_KS2 = 0.91, AUC_ROC = 0.99 and accuracy of 95.6% for Ford_A and Max_KS2 = 0.88, AUC_ROC = 0.98 and accuracy of 94.1% for Ford_B. These results have been confirmed on the competition for the noiseless data and have degraded around 15% for the noisy data.

This work presents an award winning approach for solving the NN3 forecasting competition problem. It consisted of predicting 18 future values of 111 monthly short time series. This approach consists of applying the median value of a 15-MLP ensemble for predicting each time series. The system performed very well on test data, finishing as the second best solution of the competition with a SMAPE=16.17%.

This paper proposes a hybrid differential evolutionary system (HDES) for financial time series forecasting, which performs a differential evolutionary search for the minimum dimension to determining the characteristic phase space that generates the time series phenomenon. It consists of an intelligent hybrid model composed of an artificial neural network (ANN) combined with the improved differential evolution (IDE). The proposed IDE searches for the relevant time lags for a correct time series characterization, the number of processing units in the ANN hidden layer, the ANN training algorithm and the modeling of ANN. Initially, the proposed HDES chooses the most tuned prediction model for time series representation, thus it performs a behavioral statistical test in the attempt to adjust forecast time phase distortions that appear in financial time series. An experimental analysis is conducted with the proposed HDES using two real world financial time series and five well-known performance metrics are used to assess its performance. The obtained results are compared to time-delay added evolutionary forecasting (TAEF) method.

This work introduces a new method for time series prediction - time-delay added evolutionary forecasting (TAEF) - that carries out an evolutionary search of the minimum necessary time lags embedded in the problem for determining the phase space that generates the time series. The method proposed consists of a hybrid model composed of an artificial neural network (ANN) combined with a modified genetic algorithm (GA) that is capable to evolve the complete network architecture and parameters, its training algorithm and the necessary time lags to represent the series. Initially, the TAEF method finds the most fitted predictor model and then performs a behavioral statistical test in order to adjust time phase distortions that may appear in the representation of sonic series. An experimental investigation is conducted with the method with sonic relevant time series and the results achieved are discussed and coin pared, according to several performance measures, to results found with the multilayer perteptron networks and other works reported in the literature

This paper presents a new method --- the Time-delay Added Evolutionary Forecasting (TAEF) method --- for time series prediction which performs an evolutionary search of the minimum necessary number of dimensions embedded in the problem for determining the characteristic phase space of the time series. The method proposed is inspired in F. Takens theorem and consists of an intelligent hybrid model composed of an artificial neural network (ANN) combined with a modified genetic algorithm (GA). Initially, the TAEF method finds the most fitted predictor model for representing the series and then performs a behavioral statistical test in order to adjust time phase distortions.

This work presents a new procedure for the solution of time series forecasting problems which searches for the necessary minimum quantity of dimensions embedded in the problem for determining the characteristic phase space of the phenomenon generating the time series. The proposed system is inspired in F. Takens theorem (1980) and consists of an intelligent hybrid model composed of an artificial neural network (ANN) combined with a modified genetic algorithm (GA). It is shown how this proposed model can boost the performance of time series prediction of both artificially generated time series and real world time series from the financial market. An experimental investigation is conducted with the introduced method with five different relevant time series and the results achieved are discussed and compared with previous results found in the literature, showing the robustness of the proposed approach.

In this paper, an auto-associator neural network based on the constructive cascade correlation architecture (Cascor) is investigated on an real-world signature verification problem. The traditional multilayer-perceptron trained with backpropagation is also examined in the same problem and a experimental comparison is conducted to evaluate the two network's generalization performances. The main objective is to show that constructive networks, in this case represented by the cascade-correlation, can offer in some situations a real alternative to the traditional models for the solution of practical problems. The experimental results indicate that the constructive network investigated can be efficiently applied to difficult real world pattern verification problems.

Function approximation is a very important task in environments where computation has to be based on extracting information from data samples in real world processes. Neural networks and wavenets have been seen as attractive tools for developing efficient solutions for many real world problems in function approximation. In this paper, it is shown how feedforward neural networks can be built using a different type of activation function referred to as the PPS (polynomial powers of sigmoids)-wavelet. An algorithm is presented to generate a family of PPS-wavelets that can be used to efficiently construct feedforward networks for function approximation.

In this work, an experimental evaluation of the cascade-correlation architecture is carried out in different benchmarking pattern recognition problems. An extensive experimental framework is developed to establish a comparison between the cascade-correlation network (CC) and the more traditional multilayer perceptron (MLP) and radial basis function models (RBF). The different network configurations are evaluated with respect to generalization performance in three practical real-world tasks: the diagnosis of coronary diseases (heart), the credit screening problem (card) and the recognition of handwritten characters. It is also considered the issue of catastrophic forgetting in MLP and CC models. In addition to some clear potential advantages observed in the cascade-correlation network such as the on-learning definition of the number of hidden units, the practical satisfactory results obtained suggest that the CC model may represent in some situations an alternative to other traditional models such as the MLP and RBF networks.

Summary form only given. An experimental evaluation of the neurofuzzy models NEFCLASS and FuNN is conducted in real world pattern recognition applications. The models are investigated with respect to classification performance and the number of rules generated and compared to the traditional MLP network trained with backpropagation. The models NEFCLASS and FuNN are examined in benchmarking problems from the Proben1 database and in a large-scale credit card screening problem. A comparison is established with an MLP network and the results obtained show some potential advantages of the neuro-fuzzy classifiers over the MLP particularly with respect to the ability of the neuro-fuzzy models to generate a knowledge base of rules.

Backpropagation networks are studied in the context of the recognition of handwritten and machine printed characters and, specifically, the problem of the recognition of ‘false—positive’ patterns is investigated. The idea is proposed of integrating into the processing architecture an independent set of ‘guard units’ which act as basic matchers, rejecting patterns not belonging to the training classes.