Advances in Computational Intelligence

Clustering in transaction databases can find potentially useful patterns to gain some insight into the structure of the data, which can help for effective decision-making. However, one of the critical tasks in clustering is to identify the appropriate number of clusters, which will determine the performance of any process further applied to the transaction database. This paper presents a methodology to discover the optimal structure of purchase transaction data using the Davies-Bouldin and Calinski-Harabasz validity indices to obtain the number of clusters and formed them with the farthest-first traversals algorithm. The quality of the structures previously formed is evaluated with data complexity measures such as F1, F2, F3, N1 and IR. In this work, we use the support vector machine and multi-layer perceptron classification algorithms, to determine recognition ability in classification problems of more than two classes, and in the context of separability and imbalance of classes present in the groups previously obtained. The experimental results exhibit the viability of the proposed methodology for decision-making.

The present work aims to adapt the Artificial Organic Networks (AON), a nature-inspired, supervised, metaheuristic, machine learning class, for computational finance purposes, applied as an efficient stock market index forecasting model. Thus, the proposed model aims to forecast a stock market index, with the aid of other economic indicators, employing a historic dataset of at least eleven years for all the variables. To accomplish this, a target function is proposed: a multiple non-linear regressive model. The relevance of computational finance is discussed, pointing out that is an area that has developed significantly in the last decades with different applications, some of these are: rich portfolio optimization, index-tracking, credit risk, stock investment, among others. Specifically, the Index Tracking Problem (ITP) concerns the prediction of stock market prices, being this a complex problem of the kind NP-hard. In this work, is discussed the undertaken innovative approach to implement the AON method, its main properties, as well as its implementation using the topology defined as Artificial Hydrocarbon Network (AHN), to tackle the ITP. Finally, we present the results of using a hybrid method based on K-means and the AHN configuration; within the result, the relative error obtained with this hybrid method was 0.0057.

The majority of scientific articles is available in Portable Document Format (PDF). Although PDF format has the advantage of preserving layout across platforms it does not maintain the original metadata structure, making it difficult further text processing. Despite different layouts, depending on the applied template, articles have a hierarchical structure and are divided into sections, which represent topics of specific subjects, such as methodology and results. Hence, section segmentation serves as an important step for a contextualized text processing of scientific articles. Therefore, this work applies binary classification, a supervised learning task, for section title detection in PDF scientific articles. To train the classifiers, a large dataset (more than 5 millions samples from 7,302 articles) was created through an automated feature extraction approach, comprised by 17 features, where 4 were introduced in this work. Training and testing were made for ten different classifiers for which the best F1 score reached 0.94. Finally, we evaluated our results against CERMINE, an open-source system that extracts metadata from scientific articles, having an absolute improvement in section detection of 0.19 in F1 score.

Mexican Sign Language (MSL) is the primary form of communication for the deaf community in Mexico. MSL has a different grammatical structure than Spanish; furthermore, facial expression plays a determining role in complementing context-based meaning. This turns it difficult for a hearing person without prior knowledge of the language to understand what is to be transmitted, representing an important communication barrier for deaf people. In order to face this, we present the first architecture to consider facial features as indicators of grammatical tense to develop a real-time interpreter from MSL to written Spanish. Our model uses the open source MediaPipe library to extract marks from the face, body position and hands. Three 2D convolutional neural networks are used to encode individually and extract patterns, the networks converge to a multilayer perceptron for classification. Finally, a Hidden Markov Model is used to morphosyntactically predict the most probable sequence of words based on a preloaded knowledge base. From the experiments were carried out, a precision of 94.9% was obtained with \(\sigma = 0.07\) for the recognition of 75 isolated words and 94.1% with \(\sigma = 0.09\) for the interpretation of 20 sentences in MSL in a medical context. Being an approach based on camera inputs and observing that even with a few samples an adequate generalization can be achieved, it would be feasible to scale our architecture to other sign languages and offer possibilities of efficient communication to millions of people with hearing disability.

A large number of car accidents are caused by failures in the pavement. Their automatic detection is important for pavement maintenance, however, the current public datasets of images to train and test these systems contain a few hundred samples. In this paper, we introduce a new large dataset of images with more than 2000 samples that contains the five most common risks on pavement manually annotated. We analyze and describe statistically the properties of this dataset and we establish the performance of some baseline methods in order to be useful as a benchmark. We achieve up to 89.35% accuracy in the segmentation of the different types of risk on the pavement

Acoustic scene classification is a process of characterizing and classifying the environments from sound recordings. The first step is to generate features (representations) from the recorded sound and then classify the background environments. However, different kinds of representations have dramatic effects on the accuracy of the classification. In this paper, we explored the three such representations on classification accuracy using neural networks. We investigated the spectrograms, MFCCs, and embeddings representations using different CNN networks and autoencoders. Our dataset consists of sounds from three settings of indoors and outdoors environments – thus, the dataset contains sounds from six different kinds of environments. We found that the spectrogram representation has the highest classification accuracy while MFCC has the lowest classification accuracy. We reported our findings, insights, and some guidelines to achieve better accuracy for environment classification using sounds.

Exploration is an essential part of reinforcement learning, which restricts the quality of learned policy. Hard-exploration environments are defined by huge state space and sparse rewards. In such conditions, an exhaustive exploration of the environment is often impossible, and the successful training of an agent requires a lot of interaction steps. In this paper, we propose an exploration method called Rollback-Explore (RbExplore), which utilizes the concept of the persistent Markov decision process, in which agents during training can roll back to visited states. We test our algorithm in the hard-exploration Prince of Persia game, without rewards and domain knowledge. At all used levels of the game, our agent outperforms or shows comparable results with state-of-the-art curiosity methods with knowledge-based intrinsic motivation: ICM and RND. An implementation of RbExplore can be found at https://​github.​com/​cds-mipt/​RbExplore.

Deep Learning has shown high performance in different domains, however, they need large computational resources and training data sets. Transfer learning offers an alternative to reduce both aspects but requires an effective way for selecting relevant pre-trained models for the current task. In this work, we propose a meta-learning formulation using the entropies of the feature maps produced by a pre-trained model for predicting its performance in a new target task via a regression model. Our method is tested in the time-series domain, where it obtains a better top-1 precision and a better position of a useful source task than state-of-the-art methods in 85 datasets from the UCR archive.

Few-shot learning is a relatively new technique that specializes in problems where we have little amounts of data. The goal of these methods is to classify categories that have not been seen before with just a handful of samples. Recent approaches, such as metric learning, adopt the meta-learning strategy in which we have episodic tasks conformed by support (training) data and query (test) data. Metric learning methods have demonstrated that simple models can achieve good performance by learning a similarity function to compare the support and the query data. However, the feature space learned by a given metric learning approach may not exploit the information given by a specific few-shot task. In this work, we explore the use of dimension reduction techniques as a way to find task-significant features helping to make better predictions. We measure the performance of the reduced features by assigning a score based on the intra-class and inter-class distance, and selecting a feature reduction method in which instances of different classes are far away and instances of the same class are close. This module helps to improve the accuracy performance by allowing the similarity function, given by the metric learning method, to have more discriminative features for the classification. Our method outperforms the metric learning baselines in the miniImageNet dataset by around 2% in accuracy performance.

Due to the imminent climate-change emergency, it is urgent to boost the exploitation of renewable resources to produce clean energy, being solar energy one of the most promising ones. However, one of the greatest challenges that solar energy faces is its intermittency. Thus, to get the biggest benefit from this resource, especially for photovoltaic generation, it is required to predict its availability to estimate variations in energy production. As the first step for solar radiation forecasting, a seasonality analysis is mandatory to obtain better results. In this work, we perform a seasonality analysis of solar radiation in Mexico using Machine Learning. Specifically, we accomplish a cluster analysis of solar radiation data in locations representative of the different climate conditions in Mexico to obtain a seasonality atlas of the solar resource. Cluster analysis is performed with two algorithms, k-means and k-medoids. Finally, the Silhouette method is used to validate the results.

Risk assessment is relevant in any workplace, however, there is a degree of unpredictability when dealing with flammable or hazardous materials so that detection of fire accidents by itself may not be enough. An example of this is the impingement of jet fires, where the heat fluxes of the flame could reach nearby equipment and dramatically increase the probability of a domino effect with catastrophic results. Because of this, the characterization of such fire accidents is important from a risk management point of view. One such characterization would be the segmentation of different radiation zones within the flame, so this paper presents exploratory research regarding several traditional computer vision and Deep Learning segmentation approaches to solving this specific problem. A data set of propane jet fires is used to train and evaluate the different approaches and given the difference in the distribution of the zones and background of the images, different loss functions, that seek to alleviate data imbalance, are also explored. Additionally, different metrics are correlated to a manual ranking performed by experts to make an evaluation that closely resembles the expert’s criteria. The Hausdorff Distance and Adjusted Rand Index were the metrics with the highest correlation and the best results were obtained from the UNet architecture with a Weighted Cross-Entropy Loss. These results can be used in future research to extract more geometric information from the segmentation masks or could even be implemented on other types of fire accidents.

Safety Stock is generally accepted as an appropriate inventory management strategy to deal with the uncertainty of demand and supply, as well as for limiting the risk of service loss and overproduction [6]. In particular, companies from the cosmetics and beauty industry face additional inventory management challenges derived from the strict regulatory standards applicable in different jurisdictions, in addition to the constantly changing trends, which highlight the importance of defining an accurate safety stock. In this paper, on the basis of the Linear Regression, Decision Trees, Support Vector Machine (“SVM”) and Neural Network machine learning techniques, we modeled a general Safety Stock equation and one per product category for a multinational enterprise operating in the cosmetics and beauty industry. The results of our analysis indicate that the Linear Regression is the most accurate model to generate a reasonable and effective prediction of the company’s Safety Stock.

In this paper, we introduce a parameter selection for the algorithm DBSCAN based on the \(K-neighborhood\). We change the parameters \(\epsilon \) and \(min\_points\) by a \(K-neighborhood\) (named \(\beta \)), scale, and an \(\alpha \) value. We use the scale parameter to balance the dataset. \(\beta \) is used to select \(min\_points-\epsilon \) and \(\alpha \) to reduce the value of \(min\_points\). We use homogeneity, completeness, and v-measure scores over datasets with balanced and unbalanced clusters to evaluate the performance. We compared our results against ACND and DBSCAN with the original parameter selection. Finally, we use our proposal to detect contour over 3D shapes. Our results show better performance in three of the eight datasets, and a better performance into border detection on 3D shapes.

Deep Learning has achieved great successes in various complex tasks such as image classification, detection and natural language processing. This work describes the process of designing and implementing seven deep learning approaches to perform regressions on mosquito populations from a specific region, given co-variables such as humidity, uv-index and precipitation intensity. The implemented approaches were: Recurrent Neural Networks (LSTM), an hybrid deep learning model, and a Variational Autoencoder (VAE) combined with a Multi-Layer Perceptron (MLP) which instead of using normal RGB images, uses satellite images of twelve channels from Copernicus Sentinel-2 mission. The experiments were executed on the Washington Mosquito Dataset, augmented with weather information. For this dataset, an MLP proved to achieve the best results.

Facial expressions are always manifested by people. For the human being, it is easy to recognize emotions. Technological advances applied to the recognition of expressions are growing rapidly and, in the same way, interest in research on this topic. However, at the computational level, it is a complicated task, some of the expressions of human beings are similar, for this reason, the computer can be confused at the moment of recognition. The use of machine learning models specifically artificial neural networks that have a good performance in emotions recognition is required to detect automatically feelings. This research shows the performance analysis of artificial neural networks applied to emotion datasets. The FER2013 and JAFFE datasets were used, a preprocessing of the data was carried out. For the classification, a comparison was made between artificial neural networks (Perceptron, VGG, and a Convolutional Neural Network). Optimal results were obtained in the detection of emotions.

Recent studies of brain-computer interface (BCI) have focused on the use of machine learning algorithms for the classification of brain signals. These algorithms find patterns in brain waves to distinguish between one class and another to turn them into control commands. To discuss the efficiency of classification algorithms in BCI for the classification of electroencephalogram (EEG) signals with motor images (MI), in this study twelve machine learning (ML) classifiers are applied and analyzed. The algorithms used are: 1) Convolutional network-Long short term memory (CNN-LSTM), 2) Convolutional network-gate recurrent unit (CNN-GRU), 3) Convolutional-bidirectional long short term memory (CNN-BiLSTM), 4) convolutional-bidirectional gated recurrent unit (CNN-BiGRU), 5) Random Forest, 6) Decision tree (DT), 7) Multilayer Perceptron (MLP), 8) Gaussian Naive Bayes, 9) Support Vector Machine (SVM), 10) Logistic Regression, 11) AdaBoost, 12) K-nearest neighbor (KNN). As classification tests, four mental tasks were registered, which are, the imagination of the movement of the left foot, the imagination of the movement of the left hand, state of relaxation, and mathematical activity, these mental tasks were obtained by a portable electroencephalogram (EEG) device. In the tests carried out it was found that the highest rate was 97% and the low rate was 22%.

In recent years, touchless fingerprint recognition systems have become a reliable alternative to the conventional touch fingerprint recognition system. Furthermore, with the current health crisis caused by the emergence of SARS-COV 2, the implementation of technologies that allow us to avoid direct contact with readers and devices arises as an urgent need. This article shows a system for fingerprint segmentation, filtering, and enhancement by fingerphoto technology. The dataset was acquired from smartphones on uncontrolled conditions. The proposed fingerprint recognition scheme provides an efficient preview of an automated identification system that can be extended to numerous security or administration applications. Skin model segmentation presents an accuracy of 95% over other solutions for background removal in the state of the art. For fingerphoto extraction, results were evaluated with the NIST Finger Image Quality \(NFIQ\) of the National Institute of Standards and Technology \(NIST\).

The deficit of attention on any critical activity has been a principal source of accidents leading to injuries and fatalities. Therefore the fast detection of it has to be a priority in order to achieve the safe completion of any task and also to ensure the display the maximum capabilities of the user when achieving the respective activity. While multiple methods has been developed, a new trend of non-intrusive vision based methodologies has been strongly picked by both the research and industrial communities as one with the most potential effectiveness and usability on real life scenarios.

In this paper, a new attention deficit detection system is presented. Low-weight Machine Learning algorithms will allow the use in remote applications and a variety of goal devices to avoid accidents caused by the lack of attention in complex activities. This article describes its impact, its functioning and previous work. In addition, the system is broken down into its most basic components and its results in various evaluation stages. Finally, its results in semi-real environments are presented and possible applications in real life are discussed.

The importance of urban perception computing is relatively growing in machine learning, particularly in related areas to Urban Planning and Urban Computing. This field of study focuses on developing systems to analyze and map discriminant characteristics that might directly impact the city’s perception. In other words, it seeks to identify and extract discriminant components to define the behavior of a city’s perception. This work will perform a street-level analysis to understand safety perception based on the “visual components”. As our result, we present our experimental evaluation regarding the influence and impact of those visual components on the safety criteria and further discuss how to properly choose confidence on safe or unsafe measures concerning the perceptional scores on the city street levels analysis.

Visual relocalisation is a well-known problem in the robotics community, where chromatic images are used to recognise a place that is being re-visited or re-observed again. Due to the success of deep neural networks in several computer vision tasks, convolutional neural networks have been proposed to address the visual relocalisation problem as well. However, these solutions follow the conventional off-line training in order to generate a model that can be used to regress a camera’s pose w.r.t to an input image. In this work, we present a methodology based on continual learning to address the visual relocalisation problem aiming at performing on-line model training, seeking to generate a model that is updated continuously to learn new acquired images associated with GPS coordinates. Moreover, we apply this methodology to the multi-camera case, where 8 images are acquired from a multi-rig camera, seeking to improve the localisation accuracy, this is, by using a multi-camera, we obtain a set of images observing different viewpoints of the scene for a given GPS position. Therefore, by using a voting scheme, our on-line learned model is capable of performing visual relocalisation with an accuracy of 0.78, performing at 50 fps.

Economics, social encounters, and most importantly, human life was deeply affected by the COVID-19 pandemic. The international state of contingency led vaccine manufacturers worldwide to double their efforts in developing a vaccine that would be influential in contagion rate decrease. This paper offers an overview of the worldwide vaccination process and its impact on confirmed cases. In this work, we present a time series analysis methodology to predict which country group using each of the most popular vaccines will have the less steep curve of confirmed cases in a time window of 21 days. The experiments led to 94% of the data fitting our models on average, leading to a confident suggestion on the vaccine related to the less steep foreseeable contagion slope.

This paper aims to implement time series normalization methods in order to compare situations for top countries with more deaths due to COVID-19 over the time. In this work, a dashboard set was created using Power BI for analytical dashboards, is tracked the daily data dynamics of the pandemic which is collected and represented graphically. For all data collecting were developed various web scraping scripts mainly based on bash scripting and python which extract data from specific web sites and once the initial inputs are obtained, the transforming process is started. This includes making aggregations, key performance indicators, correlations and mappings giving the facility to use that transformed data for future works. The data has been collected and treated for study from different sources [1–4]. Additionally, all the results and final data after transformations are being published on a daily basis in the following sites [5–8].

In this paper we propose a stand-alone method to identify lines of different thickness in an image exploiting the scale Space Radon Transform (SSRT) combined to multiscale image Hessian. The proposed approach does not need any prior knowledge about the image content, neither make any assumption about the image to make a decision of the presence or not of a line. This work which consists in seeking information about possible presence of linear structures in an image and exploiting this information while constructing the SSRT space, limits the SSRT computation around precomputed zones. The latter are obtained by multiscale image Hessian. As a consequence, the subsequent maxima detection is done on a restricted transform space freed from unwanted peaks that usually drown the peaks representing lines. Tests done on synthetic and real images have shown that our method highlight the useful maxima of the SSRT permitting to improve SSRT detection of lines of different thickness in an image, while preserving computation time.

Biomedical image segmentation is an important process in computer-aided diagnostic systems. Segmentation allows an image to be divided and tagged into anatomical sub-regions such as bones, muscles, blood vessels, and various pathological structures, facilitating image analysis and detecting various diseases. However, it is challenging due to the different features that images present, including noise and contrast. This article provides an overview of deep neural networks for biomedical image segmentation specifically computed tomography, dermoscopy, MRI, ultrasound, and X-ray. Additionally, best practices are discussed to improve the accuracy and sensitivity of results, including in unbalanced datasets.

Technical analysis aims to predict market movement by examining historical data through statistical procedures. Nevertheless, it is sensitive to the parameter it is working with. An optimization problem is defined to tune technical analysis parameters by minimizing an error metric for stock return prediction. Differential Evolution is a metaheuristic that provides good solutions to an optimization problem, searching for the optimal combination of parameters for technical analyzers to predict Mexican stock returns. For the application of the metaheuristic, an objective function based on a Random Forest prediction is used.

The literature has proven the use of different macroeconomic variables (MEV) to determine expected returns, such as the Capital Asset Pricing Model (CAPM) or different Arbitrage Pricing Theories (APT). This paper considers the influence of macroeconomic factors on stock prices; it is approached with a Granger-causality test on the different sector indexes of the Mexican stock exchange, to see the relationship they hold.

Instead of supervising the error from the machine learning models, it is proposed to analyze their performance in a more realistic scenario, by simulating a portfolio. Constructing a diversified portfolio is a smart way to allocate your money parting from the expected returns computed, still, other relevant factors may alter its performance.

This work shows the performance of different portfolios constructed from the same expected return computations, reaching excess returns over the benchmarks of the 12% in the 3 years analyzed.

Reference sets generated with uniformly distributed weight vectors on a unit simplex are widely used by several multi-objective evolutionary algorithms (MOEAs). They have been employed to tackle multi-objective optimization problems (MOPs) with four or more objective functions, i.e., the so-called many-objective optimization problems. These MOEAs have shown a good performance on MOPs with regular Pareto front shapes, i.e., simplex-like shapes. However, it has been observed that in many cases, their performance degrades on MOPs with irregular Pareto front shapes. In this paper, we designed a new selection mechanism that aims to promote a Pareto front shape invariant performance of MOEAs that use weight vector-based reference sets. The newly proposed selection mechanism takes advantage of weight vector-based reference sets and seven pair-potential functions. It was embedded into the non-dominated sorting genetic algorithm III (NSGA-III) to increase its performance on MOPs with different Pareto front geometries. We use the DTLZ and DTLZ\(^{-1}\) test problems to perform an empirical study about the usage of these pair-potential functions for this selection mechanism. Our experimental results show that the pair-potential functions can enhance the distribution of solutions obtained by weight vector-based MOEAs on MOPs with irregular Pareto front shapes. Also, the proposed selection mechanism permits maintaining the good performance of these MOEAs on MOPs with regular Pareto front shapes.

PSE (Parameter Sweep Experiments) applications represent a relevant class of computational applications in science, engineering and industry. These applications involve many computational tasks that are both resource-intensive and independent. For this reason, these applications are suited for Cloud environments. In this sense, Cloud autoscaling approaches are aimed to manage the execution of different kinds of applications on Cloud environments. One of the most recent approaches proposed for autoscaling PSE applications is MIA, which is based on the multi-objective evolutionary algorithm NSGA-III. We propose to endow MIA with other multi-objective optimization algorithms, to improve its performance. In this respect, we consider two well-known multi-objective optimization algorithms named SMS-EMOA and SMPSO, which have significant mechanic differences with NSGA-III. We evaluate MIA endowed with each of these algorithms, on three real-world PSE applications, considering resources available in Amazon EC2. The experimental results show that MIA endowed with each of these algorithms significantly outperforms MIA based on NSGA-III.

Scrum is an agile software development framework followed nowadays by many software companies worldwide. Since it is an iterative and incremental methodology, the software is developed in releases. For each release, the software development team and the customer agree upon a development plan. However, the context of the software project may change due to unpredicted circumstances that generally arise, for example, new software requirements or changes in the development team. Consequently, these factors force the release plan to be adjusted. When the release plan is modified, it is necessary to consider at least four criteria to minimize the economic and operational impact of these changes. Therefore, this activity can be analyzed as a multi-objective problem. In the last three decades, multi-objective evolutionary algorithms have become an effective and efficient tool to solve multi-objective problems. In this paper, we evaluate three multi-objective optimization approaches when solving the release plan rescheduling problem. Mainly, we focus our investigation on analyzing the conflict between the considered objectives and on the performance of the Pareto-based, the indicator-based, and the decomposition-based multi-objective optimization approaches.