The digitization of clinical data via electronic health records (EHRs) shows great promise for data-based decisionmaking, particularly through the analysis of unstructured clinical notes. While large language models (LLMs) offer tools for interpreting these narratives, their vast size and general-purpose design can pose difficulties in real-world healthcare settings. More specialized LLMs, tailored to healthcare contexts, are an alternative solution, although their integration into resourceconstrained environments is often problematic due to their dimensions. In this work, we present MedESM, a lightweight 1B parameter medical LLM specifically fine-tuned for structured summarization of clinical notes across twelve clinically relevant categories. Within our knowledge, MedESM is the only model at the 1B parameter scale capable of performing this task. Trained on over 40,000 real and synthetic clinical notes from the MIMIC-IV-NOTE and Asclepius datasets, MedESM leverages targeted instruction-based summarization and careful prompt design to ensure stable performance on long and complex inputs. Results show that MedESM achieves competitive quality with significantly reduced computational demands, offering a practical solution for structured documentation support in clinical settings.

This paper proposes a Machine Learning (ML) approach to perform an Ahead-of-Time (AoT) prediction of decorated particle-board production disruptions and defects. We worked with a Portuguese company that is adopting the Industry 4.0 concept aiming to improve their decorated particleboard production planning (e.g., reducing production time and waste of materials). This company’s business needs are addressed in terms of two nontrivial binary Classification tasks (production disruptions and defects). The AoT prediction is achieved by using only input attributes available before the execution of the production process. To reduce the modeling effort, we focus on Automated ML (AutoML) methods, under two main approaches: Single-Target Classification (STC) and Multi-Target Classification (MTC). The former is achieved by adopting the popular H2O AutoML tool, while the latter adopts a deep learning neural network automatically tuned by using a Bayesian search. The computational experiments adopted a realistic rolling window evaluation over recently collected industrial data (comprising 14 months). Overall, interesting predictive results were achieved by both AutoML approaches, outperforming a baseline Decision Tree method. In addition, an eXplainable Artificial Intelligence (XAI) method based on a Sensitivity Analysis (SA) was adopted, allowing the identification of the most relevant inputs, which is valuable knowledge to support the decorated particleboard production planning.

In this paper, we propose a proactive method to prevent Work-related MusculoSkeletal Disorders (WMSDs) in manufacturing industries. The integrated method includes a Motion Capture System (MCS) for data collection, a Time Series Forecasting (TSF) module using Machine Learning (ML) algorithms, a WMSD risk assessment module based on ergonomic standards, and a safety mechanism (e.g., alarm sound). We evaluated the method by analyzing shoulder abduction, rotation, and flexion movements of 12 participants working with textile machines. The computational experiments included a comparison of four ML algorithms and a baseline Naive method using a 12-fold participant cross-validation approach. Overall, the best Ahead-of-Time (AoT) TSF and WMSD risk detection empirical results were obtained by a Support Vector Machine (SVM), which required a reasonable training computational effort and provides an interesting performance for AoT TSF and high risk WMSD detection. • A proactive method is proposed to prevent Work-related MusculoSkeletal Disorders (WMSDs). • Machine Learning (ML) was used to forecast Ahead-of-Time (AoT) angular movements. • Standard ergonomics were adopted to detect upper limb high risk WMSD of 12 textile workers. • Best empirical results provided by a Support Vector Machine (SVM).

In today's fast-evolving world, the intersection of technological innovation and sustainable development has emerged as a beacon of hope for addressing global challenges. The application of data science in this context represents a powerful and transformative force, amplifying our capabilities to navigate complex societal and environmental issues. This Special Issue of Expert Systems on ‘Data Science for Sustainable Development’ is a testament to the dynamic and promising fusion of these disciplines. This last analysis examines how data science tools contribute to achieving the objectives established in September 2015 by the United Nations General Assembly, UN, within the document known as the 2030 Agenda for Sustainable Development. The objective of the Agenda is to achieve a level of growth for all countries, such as guaranteeing a sustainable future, through objectives that can be summarized in three main groups, namely the environmental, economic, and social, in the perspective of the protection of the planet. These are ambitious objectives that require adopting measures aimed at their fulfilment. The governments of the G20 were the first to represent the forerunners for the realization of this development. Given the interdisciplinary nature of data science, this can be applied to the implementation and monitoring of the achievement of the 17 objectives. The latter includes methods of collection, pre-processing, extraction of meaning/useful characteristics, methods of data exploration and predictive models. Although the results obtained are relevant, the literature on the advantages of implementing the main data science technologies in sustainability is mostly qualitative-inductive (Hoosain et al., 2020). In fact, at the moment, the empirical evidence regarding the advantages provided by data science to sustainability is reduced. However, there are numerous cases of implementation of data science applications aimed at fostering a transition towards an economy based on reuse and recovery and reducing the consumption of resources, and, above all, polluting emissions (Dantas et al., 2021; de Miranda et al., 2021; Strazzullo et al., 2023). Data science is often applied to solve business problems and brings several benefits such as process optimization, real-time monitoring, production efficiency and automation of production management. In this context, then, it is important to understand how to use data to enable measurement of the progress being made in achieving the SDGs. We have conceptualized this Special Issue in order to collect a series of articles that apply data science methodologies to the concept of the triple bottom line of sustainability considering environmental, social and economic aspects and their combinations. Indeed, the contributions in this issue encapsulate a spectrum of pioneering research, each exploring a distinctive facet of how data science intertwines with sustainable development. Through the lens of diverse scholarly articles, we delve into the symbiotic relationship between data analytics, artificial intelligence and sustainable practices. These contributions not only underscore the pressing need for innovative solutions but also illuminate the profound impact that these advancements can have on our planet and society. In the first article ‘A SCORPAN-Based Data Warehouse for Digital Soil Mapping and Association Rule Mining in Support of Sustainable Agriculture and Climate Change Analysis in the Maghreb region’, Belkadi et al. (2023) elucidate the pivotal role of data warehousing in advocating sustainable agricultural practices and discerning the ramifications of climate change in the Maghreb region. The insights garnered from this study hold immense promise for fostering resilient agricultural systems. In the second article ‘Artificial Intelligence Governance: Ethical Considerations and Implications for Social Responsibility’, Camilleri (2023) delves into the ethical dimensions of AI governance highlights the ethical imperatives entwined with the advancement of artificial intelligence, emphasizing the imperative for social responsibility in the development and deployment of AI technologies. In the third article ‘Digital Innovation and Transformation Capabilities in a Large Company’, Gupta et al. (2023) shed light on the transformative capabilities of digital innovation within large corporations underscores the necessary capabilities for achieving sustainable growth, offering invaluable insights for organizations striving towards sustainable transformations. In the fourth article ‘A Sustainable Circular Supply Chain Network Design Model for Electric Vehicle Battery Production using Internet of Things and Big Data’, Tavana et al. (2023) elucidate the integration of IoT and big data in designing a sustainable supply chain for electric vehicle battery production. This research paves the way for environmentally conscious practices in the electric vehicle industry. In the fifth article ‘Progress and Prospects of Quantum Computing in Sustainable Development: An Analytical Review’, Sood and Chauhan (2023) present a comprehensive review of quantum computing's potential applications in advancing sustainable development. Its insights open new avenues for harnessing quantum computing's potential for addressing sustainability challenges. Finally, in the sixth article ‘Edge Computing Driven Sustainable Development: A Case Study on Professional Farmer Cultivation Mechanism’, Yuan and Nie (2023) provide a meticulous case study, showcasing the applicability of edge computing in advancing sustainable practices, specifically in professional farmer cultivation mechanisms. These articles collectively underscore the versatility and relevance of data science in fostering sustainable development across multifarious domains. Each piece stands as a beacon of innovation and knowledge, offering unique perspectives that contribute to a deeper comprehension of how data science can be leveraged to confront the intricate challenges of sustainable development. The amalgamation of data science and sustainable development not only presents a convergence of disciplines but also represents an opportunity for researchers, practitioners and policymakers to collaborate and innovate for a sustainable future. In conclusion, the comprehensive body of work presented in this Special Issue on ‘Data Science for Sustainable Development’ aims to inspire and catalyse further research and initiatives in this domain. We extend our deepest appreciation to the esteemed authors for their invaluable contributions and to the rigorous reviewers for their commitment to maintaining the quality and significance of the content presented in this issue. We hope that the insights gleaned from this compilation serve as a cornerstone for shaping a future where data science and sustainable development stand as pillars for global progress and well-being. We would like to thank the authors, who contributed with their papers, the reviewers and the EXSY journal Editors.

International audience

Digital journalism has faced a dramatic change and media companies are challenged to use data science algorithms to be more competitive in a Big Data era. While this is a relatively new area of study in the media landscape, the use of machine learning and artificial intelligence has increased substantially over the last few years. In particular, the adoption of data science models for personalization and recommendation has attracted the attention of several media publishers. Following this trend, this paper presents a research literature analysis on the role of Data Science (DS) in Digital Journalism (DJ). Specifically, the aim is to present a critical literature review, synthetizing the main application areas of DS in DJ, highlighting research gaps, challenges, and opportunities for future studies. Through a systematic literature review integrating bibliometric search, text mining, and qualitative discussion, the relevant literature was identified and extensively analyzed. The review reveals an increasing use of DS methods in DJ, with almost 47% of the research being published in the last three years. An hierarchical clustering highlighted six main research domains focused on text mining, event extraction, online comment analysis, recommendation systems, automated journalism, and exploratory data analysis along with some machine learning approaches. Future research directions comprise developing models to improve personalization and engagement features, exploring recommendation algorithms, testing new automated journalism solutions, and improving paywall mechanisms.

Abstract In this paper, we propose an Intelligent Decision Support System (IDSS) for the design of new textile fabrics. The IDSS uses predictive analytics to estimate fabric properties (e.g., elasticity) and composition values (% cotton) and then prescriptive techniques to optimize the fabric design inputs that feed the predictive models (e.g., types of yarns used). Using thousands of data records from a Portuguese textile company, we compared two distinct Machine Learning (ML) predictive approaches: Single-Target Regression (STR), via an Automated ML (AutoML) tool, and Multi-target Regression, via a deep learning Artificial Neural Network. For the prescriptive analytics, we compared two Evolutionary Multi-objective Optimization (EMO) methods (NSGA-II and R-NSGA-II) when optimizing 100 new fabrics, aiming to simultaneously minimize the physical property predictive error and the distance of the optimized values when compared with the learned input space. The two EMO methods were applied to design of 100 new fabrics. Overall, the STR approach provided the best results for both prediction tasks, with Normalized Mean Absolute Error values that range from 4% (weft elasticity) to 11% (pilling) in terms of the fabric properties and a textile composition classification accuracy of 87% when adopting a small tolerance of 0.01 for predicting the percentages of six types of fibers (e.g., cotton). As for the prescriptive results, they favored the R-NSGA-II EMO method, which tends to select Pareto curves that are associated with an average 11% predictive error and 16% distance.

Under the Industry 4.0 concept, there is increased usage of data-driven analytics to enhance the production process. In particular, equipment maintenance is a key industrial area that can benefit from using Machine Learning (ML) models. In this paper, we propose a novel Remaining Useful Life (RUL) ML-based spare part prediction that considers maintenance historical records, which are commonly available in several industries and thus more easy to collect when compared with specific equipment measurement data. As a case study, we consider 18,355 RUL records from an automotive multimedia assembly company, where each RUL value is defined as the full amount of units produced within two consecutive corrective maintenance actions. Under regression modeling, two categorical input transforms and eight ML algorithms were explored by considering a realistic rolling window evaluation. The best prediction model, which adopts an Inverse Document Frequency (IDF) data transformation and the Random Forest (RF) algorithm, produced high-quality RUL prediction results under a reasonable computational effort. Moreover, we have executed an eXplainable Artificial Intelligence (XAI) approach, based on the SHapley Additive exPlanations (SHAP) method, over the selected RF model, showing its potential value to extract useful explanatory knowledge for the maintenance domain.

Nowadays, manufacturing industries still face difficulties applying traditional Work-related MusculoSkeletal Disorders (WMSDs) risk assessment methods due to the high effort required by a continuous data collection when using observational methods. An interesting solution is to adopt Inertial Measurement Units (IMUs) to automate the data collection, thus supporting occupational health professionals. In this paper, we propose a deep learning approach to predict human motion based on IMU data with the goal of preventing industrial worker problematic movements that can arise during repetitive actions. The proposed system includes an initial Madgwick filter to merge the raw inertial tri-axis sensor data into a single angle orientation time series. Then, a Machine Learning (ML) algorithm is trained with the obtained time series, allowing to build a forecasting model. The effectiveness of the developed system was validated by using an open-source dataset composed of different motions for the upper body collected in a laboratory environment, aiming to monitor the abduction/adduction angle of the arm. Firstly, distinct ML algorithms were compared for a single angle orientation time series prediction, including: three Long Short-Term Memory (LSTM) methods - a one layer, a stacked layer and a Sequence to Sequence (Seq2Seq) model; and three non deep learning methods - a Multiple Linear Regression, a Random Forest and a Support Vector Machine. The best results were provided by the Seq2Seq LSTM model, which was further evaluated for WMSD prevention by considering 11 human subject datasets and two evaluation procedures (single person and multiple person training and testing). Overall, interesting results were achieved, particularly for multiple person evaluation, where the proposed Seq2Seq LSTM has shown an excellent capability to anticipate problematic movements.

In this paper, we propose a Machine Learning (ML) approach to predict faults that may occur during the production of fabrics and that often cause production downtime delays. We worked with a textile company that produces fabrics under the Industry 4.0 concept. In particular, we deal with a client customization requisite that impacts on production planning and scheduling, where there is a crucial need of limiting machine stoppage. Thus, the prediction of machine stops enables the manufacturer to react to such situation. If a specific loom is expected to have more breaks, several measures can be taken: slower loom speed, special attention by the operator, change in the used yarn, stronger sizing recipe, etc. The goal is to model three regression tasks related with the number of weft breaks, warp breaks, and yarn bursts. To reduce the modeling effort, we adopt several Automated Machine Learning (AutoML) tools (H2O, AutoGluon, AutoKeras), allowing us to compare distinct ML approaches: using a single (one model per task) and Multi-Target Regression (MTR); and using the direct output target or a logarithm transformed one. Several experiments were held by considering Internet of Things (IoT) historical data from a Portuguese textile company. Overall, the best results for the three tasks were obtained by the single-target approach with the H2O tool using logarithm transformed data, achieving an R2 of 0.73 for weft breaks. Furthermore, a Sensitivity Analysis eXplainable Artificial Intelligence (SA XAI) approach was executed over the selected H2OAutoML model, showing its potential value to extract useful explanatory knowledge for the analyzed textile domain.

This paper presents a benchmark of supervised Automated Machine Learning (AutoML) tools. Firstly, we analyze the characteristics of eight recent open-source AutoML tools (Auto-Keras, Auto-PyTorch, Auto-Sklearn, AutoGluon, H2O AutoML, rminer, TPOT and TransmogrifAI) and describe twelve popular OpenML datasets that were used in the benchmark (divided into regression, binary and multi-class classification tasks). Then, we perform a comparison study with hundreds of computational experiments based on three scenarios: General Machine Learning (GML), Deep Learning (DL) and XGBoost (XGB). To select the best tool, we used a lexicographic approach, considering first the average prediction score for each task and then the computational effort. The best predictive results were achieved for GML, which were further compared with the best OpenML public results. Overall, the best GML AutoML tools obtained competitive results, outperforming the best OpenML models in five datasets. These results confirm the potential of the general-purpose AutoML tools to fully automate the Machine Learning (ML) algorithm selection and tuning.

Current developments on self-driving cars have increased the interest on autonomous shared taxicabs. While most self-driving technologies focus on the outside environment, there is also a need to provide in-vehicle intelligence (e.g., detect health and safety issues related with the car occupants). Set within an R&D project focused on in-vehicle cockpit intelligence, the research presented in this paper addresses an unsupervised Acoustic Anomaly Detection (AAD) task. Since data is nonexistent in this domain, we first design an in-vehicle sound event data simulator that can realistically mix background audios (recorded from car driving trips) with normal (e.g., people talking, radio on) and abnormal (e.g., people arguing, cough) event sounds, allowing the generation of three synthetic in-vehicle sound datasets. Then, we explore two main sound feature extraction methods (based on a combination of three audio features and mel frequency energy coefficients) and propose a novel Long Short-Term Memory Autoencoder (LSTM-AE) deep learning architecture for in-vehicle sound anomaly detection. Competitive results were achieved by the proposed LSTM-AE when compared with two state-of-the-art methods, namely a dense Autoencoder (AE) and a two-stage clustering.

Abstract Recently, the term “Industry 4.0” has emerged to characterize several Information Technology and Communication (ICT) adoptions in production processes (e.g., Internet‐of‐Things, implementation of digital production support information technologies). Business Analytics is often used within the Industry 4.0, thus incorporating its data intelligence (e.g., statistical analysis, predictive modelling, optimization) expert system component. In this paper, we perform a Systematic Literature Review (SLR) on the usage of Business Analytics within the Industry 4.0 concept, covering a selection of 169 papers obtained from six major scientific publication sources from 2010 to March 2020. The selected papers were first classified in three major types, namely, Practical Application, Reviews and Framework Proposal. Then, we analysed with more detail the practical application studies which were further divided into three main categories of the Gartner analytical maturity model, Descriptive Analytics, Predictive Analytics and Prescriptive Analytics. In particular, we characterized the distinct analytics studies in terms of the industry application and data context used, impact (in terms of their Technology Readiness Level) and selected data modelling method. Our SLR analysis provides a mapping of how data‐based Industry 4.0 expert systems are currently used, disclosing also research gaps and future research opportunities.

Automation and scalability are currently two of the main challenges of Machine Learning. This paper proposes an automated and distributed ML framework that automatically trains a supervised learning model and produces predictions independently of the dataset and with minimum human input. The framework was designed for the domain of telecommunications risk management, which often requires supervised learning models that need to be quickly updated by non-ML-experts and trained on vast amounts of data. Thus, the architecture assumes a distributed environment, in order to deal with big data, and Automated Machine Learning (AutoML), to select and tune the ML models. The framework includes several modules: task detection (to detect if classification or regression), data preprocessing, feature selection, model training, and deployment. In this paper, we detail the model training module. In order to select the computational technologies to be used in this module, we first analyzed the capabilities of an initial set of five modern AutoML tools: Auto-Keras, Auto-Sklearn, Auto-Weka, H2O AutoML, and TransmogrifAI. Then, we performed a benchmarking of the only two tools that address distributed ML (H2O AutoML and TransmogrifAI). Several comparison experiments were held using three real-world datasets from the telecommunications domain (churn, event forecasting, and fraud detection), allowing us to measure the computational effort and predictive capability of the AutoML tools.

International audience

Mobile performance marketing is a growing industry due to the massive adoption of smartphones and tablets. In this paper, we explore Deep Multilayer Perceptrons (MLP) to predict the Conversion Rate (CVR) of mobile users that are redirected to ad campaigns (i.e., if there will be a sale). We analyze recent real-world big data provided by a global mobile marketing company. Using a realistic rolling window validation, we conducted several experiments with different datasets (two sampling and two data traffic modes), in which we measure both the predictive binary classification performance and the computational effort. The modeling experiments include: two data preprocessing methods, the popular one-hot encoding and a proposed Percentage Categorical Pruning (PCP); and two MLP learning modes, offline (reset) and online (reuse). Overall, competitive classification results were achieved by the PCP transform and the two MLP learning modes, producing real-time predictions and comparing favorably against a Convolutional Neural Network and a Logistic Regression.

Abstract This study presents a literature analysis using a semiautomated text mining and topic modelling approach of the body of knowledge encompassed in 17 years (2000–2016) of literature published in the Wiley's Expert Systems journal, a key reference in Expert Systems (ESs) research, in a total of 488 research articles. The methodological approach included analysing countries from authors' affiliations, with results emphasizing the relevance of both U.S. and U.K. researchers, with Chinese, Turkish, and Spanish holding also a significant relevance. As a result of the sparsity found on the keywords, one of our goals became to devise a taxonomy for future submissions under 2 core dimensions: ESs' methods and ESs' applications. Finally, through topic modelling, data‐driven methods were unveiled as the most relevant, pairing with evaluation methods in its application to managerial sciences, arts, and humanities. Findings also show that most of the application domains are well represented, including health, engineering, energy, and social sciences.

[Excerpt] Expert Systems (ES) are a core element of human decision making. Initially, in the 70s and 80s, ES were focused on extracting explicit knowledge from human experts. With the availability of big data, after the 2000s, ES incorporated data-driven models, thus being associated with business intelligence, big data, data science and machine learning systems [Cortez and Santos, 2017]. The importance of data-driven models in the ES area is confirmed by the recent Wiley’s Expert Systems (EXSY) literature survey that analyzed all journal research articles published from 2000 to 2016 [Cortez et al., 2018]. The survey revealed data-driven as the most prevalent ES method type, corresponding to around 35% of all recently published EXSY papers. [...]

Abstract The discovery of knowledge through data mining provides a valuable asset for addressing decision making problems. Although a list of features may characterize a problem, it is often the case that a subset of those features may influence more a certain group of events constituting a sub‐problem within the original problem. We propose a divide‐and‐conquer strategy for data mining using both the data‐based sensitivity analysis for extracting feature relevance and expert evaluation for splitting the problem of characterizing telemarketing contacts to sell bank deposits. As a result, the call direction (inbound/outbound) was considered the most suitable candidate feature. The inbound telemarketing sub‐problem re‐evaluation led to a large increase in targeting performance, confirming the benefits of such approach and considering the importance of telemarketing for business, in particular in bank marketing.

[Excerpt] Expert Systems were proposed in the mid 1970s (Arnott & Pervan, 2014) with the goal of building computerized systems that mimic human behavior to solve real-world tasks. Such systems were based on artificial intelligence (AI) techniques, typically by adopting explicit (human understandable) knowledge, extracted from domain experts (e.g. by using interviews) and that was stored in a knowledge base (Buchanan, 1986).

Abstract The problem of Label Ranking is receiving increasing attention from several research communities. The algorithms that have been developed/adapted to treat rankings of a fixed set of labels as the target object, including several different types of decision trees (DT). One DT‐based algorithm, which has been very successful in other tasks but which has not been adapted for label ranking is the Random Forests (RF) algorithm. RFs are an ensemble learning method that combines different trees obtained using different randomization techniques. In this work, we propose an ensemble of decision trees for Label Ranking, based on Random Forests, which we refer to as Label Ranking Forests (LRF). Two different algorithms that learn DT for label ranking are used to obtain the trees. We then compare and discuss the results of LRF with standalone decision tree approaches. The results indicate that the method is highly competitive.

and leader of the Intelligent Data Systems group of Centro Algoritmi, with interests in the fields of business intelligence, data mining and learning classifier systems.He participated in several

Sentiment analysis has been increasingly applied to the stock market domain. In particular, investor sentiment indicators can be used to model and predict stock market variables. In this context, the quality of the sentiment analysis is highly dependent of the opinion lexicon adopted. However, there is a lack of lexicons adjusted to microblogging stock market data. In this work, we propose an automatic procedure for the creation of such lexicon by exploring a large set of labeled messages from StockTwits, a popular financial microblogging service, and using four statistical measures: adaptations of the known TF-IDF, Information Gain, Class Percentage, and a newly proposed Weighted Class Probability. The obtained lexicons are competitive when compared with a set of six reference lexicons. Moreover, we verified that it is beneficial to use continuous sentiment scores instead of sentiment labels.

Accurate time series forecasting is a key issue to support individual and organizational decision making. In this paper, we introduce several methods for short-term electric load forecasting. All the presented methods stem from computational intelligence techniques: Random Forest, Nonlinear Autoregressive Neural Networks, Evolutionary Support Vector Machines and Fuzzy Inductive Reasoning. The performance of the suggested methods is experimentally justified with several experiments carried out, using a set of three time series from electricity consumption in the real-world domain, on different forecasting horizons.

This paper quantifies the impact of topological characteristics on the performance of single radio multichannel IEEE 802.11 mesh networks. Topological characteristics are the number of nodes per subnetwork, the hop count, the neighbor node density, the hidden nodes, the number of nodes in the neighborhood of the gateway, and the hidden nodes in the neighborhood of the gateway. Network performance metrics are throughput, fairness and delay. The data mining Support Vector Machine (SVM) model was used to extract the relationships between the network topology metrics and the network performance metrics based on data results obtained through ns-2 simulation of random networks. The results obtained can be used as a basis to design channel assignment algorithms or to aid the deployment and management of single radio wireless mesh networks.

Abstract This article presents three methods to forecast accurately the amount of traffic in TCP/IP based networks: a novel neural network ensemble approach and two important adapted time series methods (ARIMA and Holt‐Winters). In order to assess their accuracy, several experiments were held using real‐world data from two large Internet service providers. In addition, different time scales (5 min, 1 h and 1 day) and distinct forecasting lookaheads were analysed. The experiments with the neural ensemble achieved the best results for 5 min and hourly data, while the Holt‐Winters is the best option for the daily forecasts. This research opens possibilities for the development of more efficient traffic engineering and anomaly detection tools, which will result in financial gains from better network resource management.

Evolutionary Algorithms (EAs) have been used to develop methods for Traffic Engineering (TE) over IP-based networks in the last few years, being used to reach the best set of link weights in the configuration of intra-domain routing protocols, such as OSPF. In this work, the multiobjective nature of a class of optimization problems provided by TE with Quality of Service constraints is identified. Multiobjective EAs (MOEAs) are developed to tackle these tasks and their results are compared to previous approaches using single objective EAs. The effect of distinct genetic representations within the MOEAs is also explored. The results show that the MOEAs provide more flexible solutions for network management, but are in some cases unable to reach the level of quality obtained by single objective EAs. Furthermore, a freely available software application is described that allows the use of the mentioned optimization algorithms by network administrators, in an user-friendly way by providing adequate user interfaces for the main TE tasks.

There are several supervised learning Data Mining (DM) methods, such as Neural Networks (NN), Support Vector Machines (SVM) and ensembles, that often attain high quality predictions, although the obtained models are difficult to interpret by humans. In this paper, we open these black box DM models by using a novel visualization approach that is based on a Sensitivity Analysis (SA) method. In particular, we propose a Global SA (GSA), which extends the applicability of previous SA methods (e.g. to classification tasks), and several visualization techniques (e.g. variable effect characteristic curve), for assessing input relevance and effects on the model's responses. We show the GSA capabilities by conducting several experiments, using a NN ensemble and SVM model, in both synthetic and real-world datasets.

Multi-step ahead forecasting is an important issue for organizations, often used to assist in tactical decisions. Such forecasting can be achieved by adopting time series forecasting methods, such as the classical Holt-Winters (HW) that is quite popular for seasonal series. An alternative forecasting approach comes from the use of more flexible learning algorithms, such as Neural Networks (NN) and Support Vector Machines (SVM). This paper presents a simultaneous variable (i.e. time lag) and model selection algorithm for multi-step ahead forecasting using NN and SVM. Variable selection is based on a backward algorithm that is guided by a sensitivity analysis procedure, while model selection is achieved using a grid-search. Several experiments were devised by considering eight seasonal series and the forecasts were analyzed using two error criteria (i.e. SMAPE and MSE). Overall, competitive results were achieved when comparing the SVM and NN algorithms with HW.

Unsolicited e-mail (spam) is a severe problem due to intrusion of privacy, online fraud, viruses and time spent reading unwanted messages. To solve this issue, Collaborative Filtering (CF) and Content-Based Filtering (CBF) solutions have been adopted. We propose a new CBF-CF hybrid approach called Symbiotic Data Mining (SDM), which aims at aggregating distinct local filters in order to improve filtering at a personalized level using collaboration while preserving privacy. We apply SDM to spam e-mail detection and compare it with a local CBF filter (i.e. Naive Bayes). Several experiments were conducted by using a novel corpus based on the well known Enron datasets mixed with recent spam. The results show that the symbiotic strategy is competitive in performance when compared to CBF and also more robust to contamination attacks.

The OBSERVER is a video surveillance system that detects and predicts abnormal behaviors aiming at the intelligent surveillance concept. The system acquires color images from a stationary video camera and applies state of the art algorithms to segment, track and classify moving objects. In this paper we present the behavior analysis module of the system. A novel method, called Dynamic Oriented Graph (DOG) is used to detect and predict abnormal behaviors, using real-time unsupervised learning. The DOG method characterizes observed actions by means of a structure of unidirectional connected nodes, each one defining a region in the hyperspace of attributes measured from the observed moving objects and having assigned a probability to generate an abnormal behavior. An experimental evaluation with synthetic data was held, where the DOG method outperforms the previously used N-ary Trees classifier.

OSPFOSPF is the most common intra-domain routing protocol in Wide Area Networks. Thus, optimizing OSPF weights will produce tools for traffic engineering with Quality of Service constraints, without changing the network management model. Evolutionary Algorithms (EAs) provide a valuable tool to face this NP-hard problem, allowing flexible cost functions with several metrics of the network behavior. A novel framework is proposed that enriches current models for network congestion with delay constraints, setting the basis for EAs that allocate OSPF weights, guided by a bi-objective cost function. The results show that EAs make an efficient method, outperforming common heuristics and achieving effective network behavior under unfavorable scenarios.

The forecast of Internet traffic is an important issue that has received few attention from the computer networks field. By improving this task, efficient traffic engineering and anomaly detection tools can be created, resulting in economic gains from better resource management. This paper presents a neural network ensemble (NNE) for the prediction of TCP/IP traffic using a time series forecasting (TSF) point of view. Several experiments were devised by considering real-world data from two large Internet Service Providers. In addition, different time scales (e.g. every five minutes and hourly) and forecasting horizons were analyzed. Overall, the NNE approach is competitive when compared with other TSF methods (e.g. Holt-Winters and ARIMA).

This paper describes a new approach to perform segmentation of moving objects in real-time from images acquired by a fixed color video camera and is the first tool of a major project that aspires to recognize abnormal human behavior in public areas. The moving objects detection is based on background subtraction and it is unaffected by changes in illumination, i.e., cast shadows and highlights. Furthermore it does not require a special attention during the initialization process, due to its ability to detect and rectify ghosts. The results show that with image resolutions of 380/spl times/280 at 24 bits per pixel, the time spent in the segmentation process is around 80 ms, in a 32 bits 3 GHz processor based computer.

The complex task of giving out tables to guests, according to their preferences, at a wedding party, instantiates a broader class of clustering problems, whose purpose is to group a number of entities into a number of clusters, according to a set of hard constraints, and optimizing an objective function. In order to study the application of genetic and evolutionary algorithms (GEAs) to these class of problems, some experiments were conducted. These contemplated different approaches to constraint handling, namely the use of penalty functions and decoders. The encoding issue was also studied, being compared direct and indirect representations of the problem's solutions in the chromosomes. The development of hybrid genetic operators, that combine the synergies of the GEAs paradigm with those of problem dependent heuristics, were also taken into account. The overall result is a study on the performance of several approaches to constrained optimization by GEAs, that can be used to guide the application of the paradigm in real-world problems, in the combinatorial optimization arena.