The SVM classifier often uses radial basis kernel because it has just one tunable hyperparameter, unlike polynomial kernel that has three. However, the polynomial kernel is separable and may speed up the SVM training and test, although with high degrees it is still slow because it requires many monomials. On the contrary, low degree (e.g. quadratic) polynomial kernels keep the number of monomials low even with high-dimensional inputs, being faster and extending the applicability of SVM to large scale datasets. We prove experimentally that quadratic polynomial kernel with just one hyperparameter achieves performance similar to radial basis kernel. We propose a method named increasing quadratic estimation, IQE, that calculates the hyperparameter value using only the training data, without SVM training. The proposed IQE achieves state-of-the-art performance and is very fast, because its complexity is linear on the training set size and dimensionality. The experimental work, performed on a collection of 120 classification datasets, proves that IQE: 1) outperforms and is faster than quadratic kernel without tuning; 2) is similar to radial basis and quadratic kernels tuned using grid search, being one or two orders of magnitude faster; and 3) outperforms genetic, Bayesian and particle swarm optimization, being between three and five orders of magnitude faster. Code is available from https://osf.io/nz96q Open Science Framework (OSF) .

The support vector machine (SVM) with Gaussian kernel often achieves state-of-the-art performance in classification problems, but requires the tuning of the kernel spread. Most optimization methods for spread tuning require training, being slow and not suited for large-scale datasets. We formulate an analytic expression to calculate, directly from data without iterative search, the spread minimizing the difference between Gaussian and ideal kernel matrices. The proposed direct gamma tuning (DGT) equals the performance of and is one to two orders of magnitude faster than the state-of-the art approaches on 30 small datasets. Combined with random sampling of training patterns, it also runs on large classification problems. Our method is very efficient in experiments with 20 large datasets up to 31 million of patterns, it is faster and performs significantly better than linear SVM, and it is also faster than iterative minimization. Code is available upon paper acceptance from this link: https://persoal.citius.usc.es/manuel.fernandez.delgado/papers/dgt/index.html and from CodeOcean: https://codeocean.com/capsule/4271163/tree/v1.

Abstract Comorbidity between neurodevelopmental disorders is common, especially between autism spectrum disorder (ASD) and attention deficit/hyperactivity disorder (ADHD). This study aimed to detect overlapped sensory processing alterations in a sample of children and adolescents diagnosed with both ASD and ADHD. A collection of 42 standard and 8 proposed machine learning classifiers, 22 feature selection methods and 19 unbalanced classification strategies were applied on the 6 standard question groups of the Sensory Profile-2 questionnaire. The relatively low performance achieved by state-of-the-art classifiers led us to propose the feature population sum classifier, a probabilistic method based on class and feature value populations, designed for datasets where features are discrete numeric answers to questions in a questionnaire. The proposed method achieves the best kappa and accuracy, 60% and 82.5%, respectively, reaching 68% and 86.5% combined with backward sequential feature selection, with false positive and negative rates below 15%. Since the SP2 questionnaire can be filled by parents for children from three years, our prediction can alert the clinicians with an early diagnosis in order to apply early interventions.

Dry-cured ham is a traditional Mediterranean meat product consumed throughout the world. This product is very variable in terms of composition and quality. Consumer’s acceptability of this product is influenced by different factors, in particular, visual intramuscular fat and its distribution across the slice, also known as marbling. On-line marbling assessment is of great interest for the industry for classification purposes. However, until now this assessment has been traditionally carried out by panels of experts and this methodology cannot be implement in industry. We propose a complete automatic system to predict marbling degree of dry-cured ham slices, which combines: (1) the color texture features of regions of interest (ROIs) extracted automatically for each muscle; and (2) machine learning models to predict the marbling. For the ROIs extraction algorithm more than the 90% of pixels of the ROI fall into the true muscle. The proposed system achieves a correlation of 0.92 using the support vector regression and a set of color texture features including statistics of each channel of RGB color image and Haralick’s coefficients of its gray-level version. The mean absolute error was 0.46, which is lower than the standard desviation (0.5) of the marbling scores evaluated by experts. This high accuracy in the marbling prediction for sliced dry-cured ham would allow to deploy its application in the dry-cured ham industry.

Design smell detection has proven to be a significant activity that has an aim of not only enhancing the software quality but also increasing its life cycle. This work investigates whether machine learning approaches can effectively be leveraged for software design smell detection. Additionally, this paper provides a comparatively study, focused on using balanced datasets, where it checks if avoiding dataset balancing can be of any influence on the accuracy and behavior during design smell detection. A set of experiments have been conducted-using 28 Machine Learning classifiers aimed at detecting God classes. This experiment was conducted using a dataset formed from 12,587 classes of 24 software systems, in which 1,958 classes were manually validated. Ultimately, most classifiers obtained high performances,-with Cat Boost showing a higher performance. Also, it is evident from the experiments conducted that data balancing does not have any significant influence on the accuracy of detection. This reinforces the application of machine learning in real scenarios where the data is usually imbalanced by the inherent nature of design smells. Machine learning approaches can effectively be used as a leverage for God class detection. While in this paper we have employed SMOTE technique for data balancing, it is worth noting that there exist other methods of data balancing and with other design smells. Furthermore, it is also important to note that application of those other methods may improve the results, in our experiments SMOTE did not improve God class detection. The results are not fully generalizable because only one design smell is studied with projects developed in a single programming language, and only one balancing technique is used to compare with the imbalanced case. But these results are promising for the application in real design smells detection scenarios as mentioned above and the focus on other measures, such as Kappa, ROC, and MCC, have been used in the assessment of the classifier behavior.

The support vector machine (SVM) is a very important machine learning algorithm with state-of-the-art performance on many classification problems. However, on large datasets it is very slow and requires much memory. To solve this defficiency, we propose the fast support vector classifier (FSVC) that includes: 1) an efficient closed-form training free of any numerical iterative procedure; 2) a small collection of class prototypes that avoids to store in memory an excessive number of support vectors; and 3) a fast method that selects the spread of the radial basis function kernel directly from data, without classifier execution nor iterative hyper-parameter tuning. The memory requirements of FSVC are very low, spending in average only 6$\cdot 10^{-7}$sec. per pattern, input and class, and processing datasets up to 31 millions of patterns, 30,000 inputs and 131 classes in less than 1.5 hours (less than 3 hours with only 2GB of RAM). In average, the FSVC is 10 times faster, requires 12 times less memory and achieves 4.7 percent more performance than Liblinear, that fails on the 4 largest datasets by lack of memory, being 100 times faster and achieving only 6.7 percent less performance than Libsvm. The time spent by FSVC only depends on the dataset size and thus it can be accurately estimated for new datasets, while Libsvm or Liblinear are much slower on “difficult” datasets, even if they are small. The FSVC adjusts its requirements to the available memory, classifying large datasets in computers with limited memory. Code for the proposed algorithm in the Octave scientific programming language is provided.1

The Polycystic Kidney Disease (PKD) is characterized by progressive renal cyst development and other extrarenal manifestation including Polycystic Liver Disease (PLD). Phenotypical characterization of animal models mimicking human diseases are commonly used, in order to, study new molecular mechanisms and identify new therapeutic approaches. The main biomarker of disease progression is total volume of kidney and liver in both human and mouse, which correlates with organ function. For this reason, the estimation of the number and area of the tissue occupied by cysts, is critical for the understanding of physiological mechanisms underlying the disease. In this regard, cystic index is a robust parameter commonly used to quantify the severity of the disease. To date, the vast majority of biomedical researchers use ImageJ as a software tool to estimate the cystic index by quantifying the cystic areas of histological images after thresholding. This tool has imitations of being inaccurate, largely due to incorrectly identifying non-cystic regions. We have developed a new software, named CystAnalyser (register by Universidade de Santiago de Compostela-USC, and Fundación Investigación Sanitaria de Santiago-FIDIS), that combines automatic image processing with a graphical user friendly interface that allows investigators to oversee and easily correct the image processing before quantification. CystAnalyser was able to generate a cystic profile including cystic index, number of cysts and cyst size. In order to test the CystAnalyser software, 795 cystic kidney, and liver histological images were analyzed. Using CystAnalyser there were no differences calculating cystic index automatically versus user input, except in specific circumstances where it was necessary for the user to distinguish between mildly cystic from non-cystic regions. The sensitivity and specificity of the number of cysts detected by the automatic quantification depends on the type of organ and cystic severity, with values 76.84-78.59% and 76.96-89.66% for the kidney and 87.29-93.80% and 63.42-86.07% for the liver. CystAnalyser, in addition, provides a new tool for estimating the number of cysts and a more specific measure of the cystic index than ImageJ. This study proposes CystAnalyser is a new robust and freely downloadable software tool for analyzing the severity of disease by quantifying histological images of cystic organs for routine biomedical research. CystAnalyser can be downloaded from https://citius.usc.es/transferencia/software/cystanalyser (for Windows and Linux) for research purposes.

Design Smell detection has proven to be an effective strategy to improve software quality and consequently decrease maintainability expenses.In this work, we explore the influence of the size category of the software project on the automatic detection of God Class Design Smell by different machine learning techniques.A set of experiments were conducted with eight different learning classifiers on a dataset formed by 12,588 classes of 24 systems.The results were evaluated using ROC area and Kappa tests.The classifiers change their behaviour when they are used in sets that differ in the value of the selected size information of their classes.This study concludes that it is possible to improve results, mainly in agreement, of God Class detection feeding machine learning classifiers with project size information of the classes to analyze.

Prediction of students' performance is one of the most explored issues in educational data mining. To predict if students will achieve the outcomes of the subject based on the previous results enables teachers to adapt the learning design of the subject to the teaching-learning process. However, this adaptation is even more relevant if we could predict the fulfillment of the educational objectives of a subject, since teachers should focus the adaptation on the learning resources and activities related to those educational objectives. In this paper, we present an experiment where a support vector machine is applied as a classifier that predicts if the different educational objectives of a subject are achieved or not. The inputs of the problem are the marks obtained by the students in the questionnaires related to the learning activities that students must undertake during the course. The results are very good, since the classifiers predict the achievement of the educational objectives with precision over 80%.

We present a benchmarking framework to design multi spectral systems working in the NIR range for multiple purposes. This framework is composed of a hyperspectral imaging hardware and an ad-hoc software that performs pattern recognition experiments (image acquisition, segmentation, feature extraction, feature selection, classification and evaluation steps) comparing different algorithms in every step. For each experiment, we obtain a solution using a generic hyper spectral system, but we also obtain enough data to design a specific multi-spectral system in order to decrease the overall execution time. This improvement is based in the feature se lection step, that provides the most relevant wavelengths for the problem. The framework has been tested for detecting internal and external features in potatoes, determining the origin of honey, and studying fecundity parameters in hen eggs.

Parallel perceptrons (PPs) are very simple and efficient committee machines (a single layer of perceptrons with threshold activation functions and binary outputs, and a majority voting decision scheme), which nevertheless behave as universal approximators. The parallel delta (P-Delta) rule is an effective training algorithm, which, following the ideas of statistical learning theory used by the support vector machine (SVM), raises its generalization ability by maximizing the difference between the perceptron activations for the training patterns and the activation threshold (which corresponds to the separating hyperplane). In this paper, we propose an analytical closed-form expression to calculate the PPs' weights for classification tasks. Our method, called Direct Parallel Perceptrons (DPPs), directly calculates (without iterations) the weights using the training patterns and their desired outputs, without any search or numeric function optimization. The calculated weights globally minimize an error function which simultaneously takes into account the training error and the classification margin. Given its analytical and noniterative nature, DPPs are computationally much more efficient than other related approaches (P-Delta and SVM), and its computational complexity is linear in the input dimensionality. Therefore, DPPs are very appealing, in terms of time complexity and memory consumption, and are very easy to use for high-dimensional classification tasks. On real benchmark datasets with two and multiple classes, DPPs are competitive with SVM and other approaches but they also allow online learning and, as opposed to most of them, have no tunable parameters.

In this paper we address the problem of modeling creativity in Artificial Intelligence using a Genetic or Evolutionary based approach to computing, where the universe of discourse is represented as theories or programs in an extension to the Logic Programming language, which makes possible to handle incomplete or even contradictory information in an evolutionary environment. Indeed, we present a new insight for the construction of evolutive systems that combines the potential of the knowledge representation and reasoning mechanisms, present in the logic programming languages. Here, in an evolutionary setting, the candidate solutions to model the universe of discourse are seen as evolutionary logic programs or theories, being the test whether a solution is optimal based on a measure of the quality-of-information carried by those logical theories or programs. From a point of view of the process, the quality-of-information of the universe of discourse is assessed on the fly, being therefore possible to select the best logical theory or program that models it, in terms of the same time line.

We propose a method, called Direct Kernel Perceptron (DKP), to directly calculate the weights of a single perceptron using a closed-form expression which does not require any training stage. The weigths minimize a performance measure which simultaneously takes into account the training error and the classification margin of the perceptron. The ability to learn non-linearly separable problems is provided by a kernel mapping between the input and the hidden space. Using Gaussian kernels, DKP achieves better results than the standard Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA) for a wide variety of benchmark two-class data sets. The computational cost of DKP linearly increases with the dimension of the input space and it is much lower than the corresponding to SVM.

The Parallel Perceptron (PP) is a simple neural network which has been shown to be a universal approximator, and it can be trained using the Parallel Delta (P-Delta) rule. This rule tries to maximize the distance between the perceptron activations and their decision hyperplanes in order to increase its generalization ability, following the principles of the Statistical Learning Theory. In this paper we propose a closed-form analytical expression to calculate, without iterations, the PP weights for classification tasks. The calculated weights globally optimize a cost function which takes simultaneously into account the training error and the perceptron margin, similarly to the P-Delta rule. Our approach, called Direct Parallel Perceptron (DPP) has a linear computational complexity in the number of inputs, being very interesting for high-dimensional problems. DPP is competitive with SVM and other approaches (included P-Delta) for two-class classification problems but, as opposed to most of them, the tunable parameters of DPP do not influence the results very much. Besides, the absence of an iterative training stage gives to DPP the ability of on-line learning.

This paper proposes polytope ARTMAP (PTAM), an adaptive resonance theory (ART) network for classification tasks which does not use the vigilance parameter. This feature is due to the geometry of categories in PTAM, which are irregular polytopes whose borders approximate the borders among the output predictions. During training, the categories expand only towards the input pattern without category overlap. The category expansion in PTAM is naturally limited by the other categories, and not by the category size, so the vigilance is not necessary. PTAM works in a fully automatic way for pattern classification tasks, without any parameter tuning, so it is easier to employ for nonexpert users than other classifiers. PTAM achieves lower error than the leading ART networks on a complete collection of benchmark data sets, except for noisy data, without any parameter optimization.

Palynological data are used in a wide range of applications. Some studies describe the benefits of the development of a computer system to pollinic analysis. The system should involve the detection of the pollen grains on a slice, and their classification. This paper presents a system that realizes both tasks. The latter is based on the combination of shape and texture analysis. In relation to shape parameters, different ways to understand the contours are presented. The resulting system is evaluated for the discrimination of species of the Urticaceae family which are quite similar. The performance achieved is 89% of correct pollen grain classification.

ART neural networks are important tools for online supervised pattern recognition. They use internal categories with pre-defined geometry, given by the category choice function. Pre-defined geometry limits the ability of the categories to fit complex borders among output predictions for a given data set, and may contribute to the category proliferation problem. This work proposes Polytope ARTMAP (PTAM), whose category representation regions have general geometry-polytopes in R/sup n/ whose vertices are selected training patterns. The category borders compose a piece-wise linear approximation to the borders among predictions. Overlapping among categories is avoided in PTAM because they do not need to overlap in order to keep their geometry during learning. The choice function does not depend on the category size. Category growing is only limited by the other categories, and the vigilance parameter can be removed, so that PTAM learns a training data set without any parameter tuning.

This paper describes MART, an ART-based neural network for adaptive classification of multichannel signal patterns without prior supervised learning. Like other ART-based classifiers, MART is especially suitable for situations in which not even the number of pattern categories to be distinguished is known a priori; its novelty lies in its truly multichannel orientation, especially its ability to quantify and take into account during pattern classification the different changing reliability of the individual signal channels. The extent to which this ability can reduce the creation of spurious or duplicate categories (a major problem for ART-based classifiers of noisy signals) is illustrated by evaluation of its performance in classifying QRS complexes in two-channel ECG traces which were taken from the MIT-BIH database and contaminated with noise.