Artificial Intelligence Applications and Innovations

In this paper counterfactual thinking is addressed based on literature mainly from Neuroscience and Psychology. A detailed literature review was conducted in identifying processes, neural correlates and theories related to counterfactual thinking from different disciplines. A familiar scenario with respect to counterfactual thinking was identified. Based on the literature, an adaptive self-modeling network model was designed. This model captures the complex process of counterfactual thinking and the learning and control involved.

Mental disorders are more and more seen as based on complex networks of symptoms and predispositions that create the disorder as an emergent behaviour of the network’s dynamics. This paper aims to provide a computational model reflecting the adaptive causal relations between anxiety, stress and physical exercise based on a network-oriented modelling approach. The model was evaluated by executing several simulations and validated through an examination of its emergent properties and their cross-reference to the available literature. The created model offers the possibility of simulating different treatments, and offers a basis to develop a virtual patient model.

Data availability is important when researchers want to apply artificial intelligence algorithms to extract biomarkers and generate predictive models for disease diagnosis, response to treatment and prognosis. For cutaneous melanoma clinical, biological and imaging data are scattered through the web. ebioMelDB is the first database to integrate the widest collections of RNA-Seq gene expression and clinical data with clinical and dermoscopy images, all manually curated and organized in categories. ebioMelDB aspires also to host our under development predictive models in cutaneous melanoma diagnosis, response to treatment and prognosis based on combinations of the different data types hosted. As a first step towards this direction, we apply an ensemble dimensionality reduction technique employing a multi-objective optimization heuristic algorithm that finds the best feature subset, the best classifier among linear SVM, Radial Basis Function Kernel SVM and random forest and their optimal parameters to predict the vital status of patients in different time windows based on a large cohort of patients’ gene expression data. The results are very encouraging in performance metrics compared with state-of-the-art algorithms. The database is available at http://www.med.upatras.gr/ebioMelDB .

Biomedical Named Entity Recognition is a difficult task, aimed to identify all named entities in medical literature. The importance of the task becomes apparent as these entities are used to identify key features, enable better search results and can accelerate the process of reviewing related evidence to a medical case. This practice is known as Evidence-Based Medicine (EBM) and is globally used by medical practitioners who do not have the time to read all the latest developments in their respective fields. In this paper we propose a methodology which achieves state-of-the-art results in a plethora of Biomedical Named Entity Recognition datasets, with a lightweight approach that requires minimal training. Our model is end-to-end and capable of efficiently modeling significantly longer sequences than previous models, benefiting from inter-sentence dependencies.

In recent years, Electrocardiogram (ECG) applications are blooming, such as cardiovascular disease detection and mental condition assessment. To protect the sensitive ECG data from data breach, ECG biometrics system are proposed. Compared to the traditional biometric systems, ECG biometric is known to be ubiquitous, difficult to counterfeit and more suitable in cleanroom or IC fabs. ECG biometric system mainly contains identification task and verification task, and Deep-ECG is the state-of-the-art work in both tasks. However, Deep-ECG only trained on one specific dataset, which ignored the intra-variability of different ECG signals across different situations. Moreover, Deep-ECG used cross-entropy loss to train the deep convolutional neural networks (CNN) model, which is not the most appropriate loss function for such embedding-based problem. In this paper, to solve the above problems, we proposed a scalable NPairLoss-based Deep-ECG (SNL-Deep-ECG) system for ECG verification on a hybrid dataset, mixed with four public ECG datasets. We modify the preprocessing method and trained the deep CNN model with NPairLoss. Compared with Deep-ECG, SNL-Deep-ECG can reduce 90% of the signal collection time during inference with only 0.9% AUC dropped. Moreover, SNL-Deep-ECG outperforms Deep-ECG for approximately 3.5% Area Under ROC Curve (AUC) score in the hybrid dataset. Moreover, SNL-Deep-ECG can maintain its verification performance over the increasing number of the subjects, and thus to be scalable in terms of subject number. The final performance of the proposed SNL-Deep-ECG is 0.975/0.970 AUC score on the seen/unseen-subject task.

Microarray data collects information from tissues that could be used in early diagnosis such as cancer. However, the classification of microarray data is a challenging task due to the high number of features and a small number of samples leading to poor classification accuracy. Feature selection is very effective in reducing dimensionality; it eliminates redundant and irrelevant features to enhance the classifier’s performance. In order to shed light on the strengths and weaknesses of the existing techniques, we compare the performances of five embedded feature selection methods namely decision tree, random forest, lasso, ridge, and SVM-RFE. Ten well-known microarray datasets are tested. Obtained results show the outperformance of SVM-RFE in term of accuracy, and comes in the second position after decision tree in terms of number of selected features and execution time.

Artificial Intelligence brings exciting innovations in all aspects of life and creates new opportunities across industry sectors. At the same time, it raises significant questions in terms of trust, ethics, and accountability. This paper offers an introduction to the AI4Media project, which aims to build on recent advances of AI in order to offer innovative tools to the media sector. AI4Media unifies the fragmented landscape of media-related AI technologies by investigating new learning paradigms and distributed AI, exploring issues of AI explainability, robustness and privacy, examining AI techniques for content analysis, and exploiting AI to address major societal challenges. In this paper, we focus on our vision of how such AI technologies can reshape the media sector, by discussing seven industrial use cases that range from combating disinformation in social media and supporting journalists for news story creation, to high quality video production, game design, and artistic co-creation. For each of these use cases, we highlight the present challenges and needs, and explain how they can be efficiently addressed by using innovative AI-driven solutions.

The research problem that is the interest in this thesis is to understand the Big Data Analytics (BDA) potential in achieving a much better Supply Chain Management (SCM). Based on this premise, it was conducted a Regression Predictive Model to comprehend the usage of Big Data Analytics in SCM and to have insights of the requirements for the potential applications of BDA. In this study were analyzed the main sources of BDA utilized in present by Supply Chain professionals and it was provided future suggestions. The findings of the study suggest that BDA may bring operational and strategic benefit to SCM, and the application of BDA may have positive implication for industry sector.

This paper address a relevant business analytics need of a chemical company, which is adopting an Industry 4.0 transformation. In this company, quality tests are executed at the Analytical Laboratories (AL), which receive production samples and execute several instrumental analyses. In order to improve the AL stock warehouse management, a Machine Learning (ML) project was developed, aiming to estimate the AL materials consumption based on week plans of sample analyses. Following the CRoss-Industry Standard Process for Data Mining (CRISP-DM) methodology, several iterations were executed, in which three input variable selection strategies and two sets of AL materials (top 10 and all consumed materials) were tested. To reduce the modeling effort, an Automated Machine Learning (AutoML) was adopted, allowing to automatically set the best ML model among six distinct regression algorithms. Using real data from the chemical company and a realistic rolling window evaluation, several ML train and test iterations were executed. The AutoML results were compared with two time series forecasting methods, the ARIMA methodology and a deep learning Long Short-Term Memory (LSTM) model. Overall, competitive results were achieved by the best AutoML models, particularly for the top 10 set of materials.

Automated machine learning (AutoML) supports ML engineers and data scientists by automating tasks like model selection and hyperparameter optimization. A number of AutoML solutions have been developed, open-source and commercial. We propose a concept called OMA-ML (Ontology-based Meta AutoML) that combines the strengths of existing AutoML solutions by integrating them (meta AutoML).OMA-ML is based on a ML ontology that guides the meta AutoML process. It supports multiple user groups, with and without programming skills. By combining the strengths of AutoML solutions, it supports any number of ML tasks and ML libraries.

Solving partitioning problems in random environments is a classic and challenging task, and has numerous applications. The existing Object Migration Automaton (OMA) and its proposed enhancements, which include the Pursuit and Transitivity phenomena, can solve problems with equi-sized partitions. Currently, these solutions also include one where the partition sizes possess a Greatest Common Divisor (GCD). In this paper, we propose an OMA-based solution that can solve problems with both equally and non-equally-sized groups, without restrictions on their sizes. More specifically, our proposed approach, referred to as the Partition Size Required OMA (PSR-OMA), can solve general partitioning problems, with the only additional requirement being that the unconstrained partitions’ sizes are known a priori. The scheme is a fundamental contribution in the field of partitioning algorithms, and the numerical results presented demonstrate that PSR-OMA can solve both equi-partitioning and non-equi-partitioning problems efficiently, and is the only known solution that resolves this problem.

Autonomous Vehicles (AVs) can potentially reduce the accident risk while a human is driving. They can also improve the public transportation by connecting city centers with main mass transit systems. The development of technologies that can provide a sense of security to the passenger when the driver is missing remains a challenging task. Moreover, such technologies are forced to adopt to the new reality formed by the COVID-19 pandemic, as it has created significant restrictions to passenger mobility through public transportation. In this work, an image-based approach, supported by novel AI algorithms, is proposed as a service to increase autonomy of non-fully autonomous people such as kids, grandparents and disabled people. The proposed real-time service, can identify family members via facial characteristics and efficiently ignore face masks, while providing notifications for their condition to their supervisor relatives. The envisioned AI-supported security framework, apart from enhancing the trust to autonomous mobility, could be advantageous in other applications also related to domestic security and defense.

Real-word intelligent agents must be able to detect sudden and unexpected changes to their task environment and effectively respond to those changes in order to function properly in the long term. We thus isolate a set of perturbations that agents ought to address and demonstrate how task-agnostic perturbation detection and mitigation mechanisms can be integrated into a cognitive robotic architecture. We present results from experimental evaluations of perturbation mitigation strategies in a multi-robot system that show how intelligent systems can achieve higher levels of autonomy by explicitly handling perturbations.

Data available today in smart monitoring applications such as smart buildings, machine health monitoring, smart healthcare, etc., is not centralized and usually supplied by a number of different devices (sensors, mobile devices and edge nodes). Due to which the data has a heterogeneous nature and provides different perspectives (views) about the studied phenomenon. This makes the monitoring task very challenging, requiring machine learning and data mining models that are not only able to continuously integrate and analyze multi-view streaming data, but also are capable of adapting to concept drift scenarios of newly arriving data. This study presents a multi-view clustering approach that can be applied for monitoring and analysis of streaming data scenarios. The approach allows for parallel monitoring of the individual view clustering models and mining view correlations in the integrated (global) clustering models. The global model built at each data chunk is a formal concept lattice generated by a formal context consisting of closed patterns representing the most typical correlations among the views. The proposed approach is evaluated on two different data sets. The obtained results demonstrate that it is suitable for modelling and monitoring multi-view streaming phenomena by providing means for continuous analysis and pattern mining.

A significant challenge for the growing world of data is to analyze, classify and manipulate spatial data. The challenge starts with the clustering process, which can be defined to characterize the spatial data with their relative properties in different groups or classes. This process can be performed using many different methods like grids, density, hierarchical and others. Among all these methods, the use of density for grouping leads to a lower noise data in result, which is called Density-Based Spatial Clustering of Applications with Noise (DBSCAN). The DBSCAN algorithm defines a data set in a group and separates the group from the other groups based on the density of the data surrounding the selection of data points. These data points and the density of the data are calculated depending on two parameters. One parameter is used as the radius of the data point to find the neighborhood data points. Another parameter is used to identify the noise in the collected data by keeping the minimum number of data points for the data density. Like other popular method k-means, DBSCAN does not require any input of the cluster number. It can sort the data set with the number of clusters according to data density. The purpose of this article is to explain the Efficient Density-based Spatial Clustering of Applications with Noise (DBSCAN) using a sample of data set, compare the results, identify the constraints, and suggest some possible solutions.

The process of surface coating is widely applied in the manufacturing industry. The accuracy of coating strongly affects the mechanical properties of the coated components. This work suggests the use of Self-Organizing Maps (Kohonen neural networks) for an optimal robotic beam trajectory planning for surface coating applications. The trajectory is defined by the one-dimensional sequence of neurons around a triangulated substrate and the neuron weights are defined as the position, beam vector and node velocity. During the training phase, random triangles are selected according to local curvature and the weights of the neurons whose beam coats the selected triangles are gradually adapted. This is achieved using a complicated coating thickness model as a function of stand-off distance, spray impact angle and beam surface spot speed. Initial results are presented from three objects widely used in manufacturing. The accuracy of this method is validated by comparing the simulated coating resulting from the SOM-planned trajectory to the coating performed for the same objects by an expert.

Multi-Frame classification applications are constituted by instances composed by a package of image frames, such as videos, which frequently require very high computational re-sources. Furthermore, when the input instances contain a large proportion of noise, then the incorporation of noise filtering pre-processing techniques are considered essential. In this work, we propose an AutoEncoder Convolutional Neural Network model for Multi-Frame input applications. The AutoEncoder model aims to reduce the huge dimensional size of the initial instances, compress useful information while simultaneously remove the noise from each frame. Finally, a Convolutional Neural Network classification model is applied on the new transformed and compressed data instances. As a case study scenario for the proposed framework, we utilize Ultrasound images (image slices/frames extracted from every patient via a portable ultrasound device) for Sarcopenia detection. Based on our experimental re-sults the proposed framework outperforms traditional approaches.

X-ray computed tomography (XCT) is an established non-destructive testing (NDT) method that, in combination with automatic evaluation routines, can be successfully used to establish a reliable 100% inline inspection system for defect detection of cast parts. While these systems are robust in automatically localizing suspected defects, human know-how in a secondary assessment and decision-making step remains indispensable to avoid an excess of rejected parts. Rather than changing the existing defect detection system and risking difficult to anticipate changes to a solid evaluation process, we propose the integration of human know-how in a subsequent support system through end-to-end learning. Using XCT data and the corresponding decisions performed by the XCT operator, we aim to support and possibly automate the secondary quality assessment process. In our paper we present a Convolutional Neural Network (CNN) architecture to predict both, the final decision of the XCT operator and a defect class indication, for cast parts rejected by the defect detection system based on XCT slice images. On a dataset of 19,459 defect records categorized in 7 classes, we achieved an accuracy of 92% for the decision and 93% for the defect class indication on the testing split. We further show that, by binding decisions to the reliability of the predicted defect class, our model has the potential to enhance also a production process with a near-faultless condition. Based on production-line data, we estimate that our model can reliably relabel 11% of defects reported during production and provide a defect class indication for another 57%.

In this paper we exploit cross-lingual models to enable dialogue act recognition for specific tasks with a small number of annotations. We design a transfer learning approach for dialogue act recognition and validate it on two different target languages and domains. We compute dialogue turn embeddings with both a CNN and multi-head self-attention model and show that the best results are obtained by combining all sources of transferred information. We further demonstrate that the proposed methods significantly outperform related cross-lingual DA recognition approaches.

The just-in-time estimation of farmland traits such as biomass yield can aid considerably in the optimisation of agricultural processes. Data in domains such as precision farming is however notoriously expensive to collect and deep learning driven modelling approaches need to maximise performance but also acknowledge this reality. In this paper we present a study in which a platform was deployed to collect data from a heterogeneous collection of sensor types including visual, NIR, and LiDAR sources to estimate key pastureland traits. In addition to introducing the study itself we address two key research questions. The first of these was the trade off of multi-modal modelling against a more basic image driven methodology, while the second was the investigation of patch size variability in the image processing backbone. This second question relates to the fact that individual images of vegetation and in particular grassland are texturally rich, but can be uniform, enabling subdivision into patches. However, there may be a trade-off between patch-size and number of patches generated. Our modelling used a number of CNN architectural variations built on top of Inception Resnet V2, MobileNet, and shallower custom networks. Using minimum Mean Absolute Percentage Error (MAPE) on the validation set as our metric, we demonstrate strongest performance of 28.23% MAPE on a hybrid model. A deeper dive into our analysis demonstrated that working with fewer but larger patches of data performs as well or better for true deep models – hence requiring the consumption of less resources during training.

In the last years, AI systems, in particular neural networks, have seen a tremendous increase in performance, and they are now used in a broad range of applications. Unlike classical symbolic AI systems, neural networks are trained using large data sets and their inner structure containing possibly billions of parameters does not lend itself to human interpretation. As a consequence, it is so far not feasible to provide broad guarantees for the correct behaviour of neural networks during operation if they process input data that significantly differ from those seen during training. However, many applications of AI systems are security- or safety-critical, and hence require obtaining statements on the robustness of the systems when facing unexpected events, whether they occur naturally or are induced by an attacker in a targeted way. As a step towards developing robust AI systems for such applications, this paper presents how the robustness of AI systems can be practically examined and which methods and metrics can be used to do so. The robustness testing methodology is described and analysed for the example use case of traffic sign recognition in autonomous driving.

Bibliographic coupling and co-citation analysis methodologies were proposed in the early 60’s and 70’s to study the structure and production of scientific communities. Bibliographic coupling is fundamental to understand the current state of a particular research area and its possible and potential future direction. Co-citation analysis, instead, is used to map the roots of academic works, fundamental to the development of a specific research field. With the first method, papers which have a common reference are grouped and the strength of their link results from the number of references in common. The second, instead, groups together the papers co-cited by one or more documents. Both methodologies assume that the papers citing the same articles or cited from the same article have similar aspects. Since until now these two methodologies have been considered separately, a new algorithm, based on bicluster analysis, which applies them together, is proposed. The importance of this new methodology is therefore to group together the paired bibliographic papers and the co-cited references but keeping them divided. In the resulting bicluster, the references grouped together represent the roots from which was born the trend to which the citing papers, grouped together, adhere.

Word clouds are a very useful tool for summarizing textual information. They can be used to illustrate the most frequent and important words of text documents or a set of text documents. In that respect they can also be used for topic visualisation. In this paper we present an experiment investigating how the crowd understands topics visualised via word clouds. In the experiment we use the topics mined from Instagram hashtags of a set of Instagram images corresponding to 30 different subjects. By subject we mean the research hashtag we use to gather pairs of Instagram images and hashtags. With the aid of an innovative topic modelling method, developed in a previous work, we constructed word clouds for the visualisation of each topic. Then we used a popular crowdsourcing platform (Appen) to let users identify the topic they believe each word cloud represents. The results show some interesting variations across subjects which are analysed and discussed in detail throughout the paper. Given that the topics were mined from Instagram hashtags, the current study provides useful insights regarding the appropriateness of hashstags as image annotation tags.

Automated trading is an approach to investing whereby market predictions are combined with algorithmic decision-making strategies for the purpose of generating high returns while minimizing downsides and risk. Recent advancements in Machine and Deep learning algorithms has led to new and sophisticated models to improve this functionality. In this paper, a comparative analysis is conducted concerning eight studies which focus on the American and the European stock markets. The simple method of Golden Cross trading strategy is being utilized for the assessment of models in real-world trading scenarios. Backtesting was performed in two indices, the S&P 500 and the EUROSTOXX 50, resulting in relative good performance, aside from the significant downfall in global markets due to COVID-19 outbreak, which appeared to affect all models.

Differentiable rendering is a very successful technique that applies to a Single-View 3D Reconstruction. Current renderers use losses based on pixels between a rendered image of some 3D reconstructed object and ground-truth images from given matched viewpoints to optimise parameters of the 3D shape.These models require a rendering step, along with visibility handling and evaluation of the shading model. The main goal of this paper is to demonstrate that we can avoid these steps and still get reconstruction results as other state-of-the-art models that are equal or even better than existing category-specific reconstruction methods. First, we use the same CNN architecture for the prediction of a point cloud shape and pose prediction like the one used by Insafutdinov & Dosovitskiy. Secondly, we propose the novel effective loss function that evaluates how well the projections of reconstructed 3D point clouds cover the ground-truth object’s silhouette. Then we use Poisson Surface Reconstruction to transform the reconstructed point cloud into a 3D mesh. Finally, we perform a GAN-based texture mapping on a particular 3D mesh and produce a textured 3D mesh from a single 2D image. We evaluate our method on different datasets (including ShapeNet, CUB-200-2011, and Pascal3D+) and achieve state-of-the-art results, outperforming all the other supervised and unsupervised methods and 3D representations, all in terms of performance, accuracy, and training time.

The use of smartwatches as devices for tracking one’s health and well-being is becoming a common practice. This paper demonstrates the feasibility of running a real-time personalized deep learning-based fall detection system on a smartwatch device using a collaborative edge-cloud framework. In particular, we demonstrate how we automate the fall detection pipeline, design an appropriate UI on the small screen of the watch, and implement strategies for the continuous data collection and automation of the personalization process with the limited computational and storage resources of a smartwatch.

Recently, there have been advances in using unsupervised learning methods for Acoustic Anomaly Detection (AAD). In this paper, we propose an improved version of two deep AutoEncoders (AE) for unsupervised AAD for six types of working machines, namely Dense and Convolutional AEs. A large set of computational experiments was held, showing that the two proposed deep autoencoders, when combined with a mel-spectrogram sound preprocessing, are quite competitive and outperform a recently proposed AE baseline. Overall, a high-quality class discrimination level was achieved, ranging from 72% to 92%.

Employing convolutional neural network models for large scale datasets represents a big challenge. Especially embedded devices with limited resources cannot run most state-of-the-art model architectures in real-time, necessary for many applications. This paper proves the applicability of shunt connections on large scale datasets and narrows this computational gap. Shunt connections is a proposed method for MobileNet compression. We are the first to provide results of shunt connections for the MobileNetV3 model and for segmentation tasks on the Cityscapes dataset, using the DeeplabV3 architecture, on which we achieve compression by 28%, while observing a 3.52 drop in mIoU. The training of shunt-inserted models are optimized through knowledge distillation. The full code used for this work will be available online.

In the last few decades, as industrial control systems (ICSs) became more interconnected via modern networking techniques, there has been a growing need for new security and monitoring techniques to protect these systems. Advanced cyber-attacks on industrial systems take multiple steps to reach ICS end devices. However, current anomaly detection systems can only detect attacks on individual local devices, and they do not consider the impact or consequences of an individual attack on the rest of the ICS devices. In this paper, we aim to explore how deep learning recurrent neural networks and correlation analysis techniques can be used collaboratively for anomaly detection in an ICS network on the scale of the entire systems. For each detected attack, our presented system-wide anomaly detection method will predict the next step of the attack. We use iTrust SWaT dataset and Power System Attack datasets from MSU national Labs to explore how the addition of correlation analysis to recurrent networks can expand anomaly detection methods to the system-wide scale.

The benefits of deep neural networks (DNNs) have become of interest for safety critical applications like medical ones or automated driving. Here, however, quantitative insights into the DNN inner representations are mandatory [10]. One approach to this is concept analysis, which aims to establish a mapping between the internal representation of a DNN and intuitive semantic concepts. Such can be sub-objects like human body parts that are valuable for validation of pedestrian detection. To our knowledge, concept analysis has not yet been applied to large object detectors, specifically not for sub-parts. Therefore, this work first suggests a substantially improved version of the Net2Vec approach [5] for post-hoc segmentation of sub-objects. Its practical applicability is then demonstrated on a new concept dataset by two exemplary assessments of three standard networks, including the larger Mask R-CNN model [9]: (1) the consistency of body part similarity, and (2) the invariance of internal representations of body parts with respect to the size in pixels of the depicted person. The findings show that the representation of body parts is mostly size invariant, which may suggest an early intelligent fusion of information in different size categories.

Technology has had a remarkable influence on music. As society advances technologically, the music industry does as well. An example that illustrates the use of technology in music is the use of artificial intelligence (AI) as a creative and inspiring tool. Music helps shape emotional responses, creates a rhythm, and comments on the action. It is often a very crucial element to any experience. However, music, like any form of art, is an extremely challenging field to tackle using AI. The amount of information in a musical structure can be overwhelmingly large. If we factor in the different and unpredictable nuances invoked by human imperfection and emotion, it becomes clear why, even though AI excels at handling large amounts of data, generating good music can be very challenging, especially when it comes to Jazz and similarly complex genres.

Nowadays Internet of Things (IoT) is a major paradigm shift that will mark an epoch in communication technology such that every physical object can be connected to the Internet. 5G makes a significant breakthrough in the traditional mobile communication system and support the applications of IoT in various fields including business, manufacturing, health care and transportation.5G is increasing the service capability of future IoT and operates and connects the whole society. 5G is facing enormous challenges when it supports differentiated applications with a uniform technical framework. In recent years, Artificial Intelligence (AI) is rising to these challenges with the rapid development. It is a potential solution to the problems in the 5G era and will lead to a revolution in the capabilities and concepts of the communication systems. Many researches have already been done for applying AI in 5G. In this paper, we focus on clarifying the promising research directions with the greatest potential rather than trying to review all the existing literatures. In this research, 5G can be anticipated to achieve significantly better performance and more convenient implementations compared to the traditional communication systems. With the inspiring research paradigms introduced in this paper, we are looking forward to the remarkable achievements of AI in 5G in the near future.

Forecasting time series is an important problem addressed for years. Despite that, it still raises an active interest of researchers. The main issue related to that problem is the inherent uncertainty in data which is hard to be represented in the form of a forecasting model. To solve that issue, a fuzzy model of time series was proposed. Recent developments of that model extend the level of uncertainty involved in data using intuitionistic fuzzy sets. It is, however, worth noting that additional fuzziness exhibits nonlinear behavior. To cope with that issue, we propose a time series model that represents both high uncertainty and non-linearity involved in the data. Specifically, we propose a forecasting model integrating intuitionistic fuzzy sets with neural networks for predicting metal prices. We validate our approach using five financial multivariate time series. The results are compared with those produced by state-of-the-art fuzzy time series models. Thus, we provide solid evidence of high effectiveness of our approach for both one- and five-day-ahead forecasting horizons.

Brain-inspired Hyperdimensional (HD) computing is an emerging technique for low-power/energy designs in many machine learning tasks. Recent works further exploit the low-cost quantized (bipolarized or ternarized) HD model and report dramatic improvements in energy efficiency. However, the quantization loss of HD models leads to a severe drop in classification accuracy. This paper proposes a projection-based quantization framework for HD computing (PQ-HDC) to achieve a flexible and efficient trade-off between accuracy and efficiency. While previous works exploit thresholding-quantization schemes, the proposed PQ-HDC progressively reduces quantization loss using a linear combination of bipolarized HD models. Furthermore, PQ-HDC allows quantization with flexible bit-width while preserving the computational efficiency of the Hamming distance computation. Experimental results on the benchmark dataset demonstrate that PQ-HDC achieves a 2.82% improvement in accuracy over the state-of-the-art method.

Brain-inspired Hyperdimensional Computing (HDC), a machine learning (ML) model featuring high energy efficiency and fast adaptability, provides a promising solution to many real-world tasks on resource-limited devices. This paper introduces an HDC-based user adaptation framework, which requires efficient fine-tuning of HDC models to boost accuracy. Specifically, we propose two techniques for HDC, including the learnable projection and the fusion mechanism for the Associative Memory (AM). Compared with the user adaptation framework based on the original HDC, our proposed framework shows 4.8% and 3.5% of accuracy improvements on two benchmark datasets, including the ISOLET dataset and the UCIHAR dataset, respectively.

The paper presents the concept and demonstrator of “SAVE – SAfety of elderly people and Vicinity Ensuring”, an integrated approach of personal and residential security. This AAL (Active Assisted Living) system aims to support elderly end-users staying in their familiar home and surroundings for as long as possible, being safely and permanently in contact with their caregivers. The top-down Universal Modeling Language (UML) service orientation enabled a unified co-design of both human and machine communications. Furthermore, the present concept includes an integrated subscription mechanism for people and devices, in a multi-tier user structure and a central-local distributed processing (Cloud-Edge). There are presented the end-user perspective, selection and involvement (in pilot prototyping of the SAVE demonstrator), considering cognitive age-related issues, oriented on usability and on the functional specifications of interfacing via web-apps (independent of the mobile platforms operating system). Location Based Services (LBS) manage data from the Geographical Information Systems (GIS) in a unified modern way, based on LISP – Location - (from) Identity Separation Protocol. These services of “orientation” – localization, real-time tracking – are integrated with another aspect of “restoring the referential” (both personal and communitarian): (re-) planning service, improving alert information in case of emergency, bilateral push-pull of security notifications. Restoring the end-users well-being takes benefit of eHealth and actigraphy services. The other approach, bottom-up, is based on micro-services software/netware implementation, modern smartwatches/smartphones and wearable devices, leveraging their programmability and multi-modal and multi-range (near-field or cellular) communications.

A Building Energy Management System (BEMS) is a fundamental computer-based system aiming at optimizing management of building assets towards energy savings without compromising occupants’ comfort, while developing a potential to apply demand response strategies. Building energy management systems nowadays have evolved and comprise of heterogeneous components and complex architectures. Building management systems monitor indoor building conditions while controlling building assets like lights, security systems, and heating and air ventilation’s systems. Furthermore, to facilitate the specification of the building’s attributes for building energy management systems the use of semantic technologies is needed within building energy tools. This paper undertakes a comprehensive review of Building Energy systems in the context of Internet of Energy.

Mislabeled data in large datasets can quickly degrade the performance of machine learning models. There is a substantial base of work on how to identify and correct instances in data with incorrect annotations. However, time series data pose unique challenges that often are not accounted for in label noise detecting platforms. This paper reviews the body of literature concerning label noise and methods of dealing with it, with a focus on applicability to time series data. Time series data visualization and feature extraction techniques used in the denoising process are also discussed.

Learner Management Systems (LMSs) are widely deployed across the industry as they provide a cost-saving approach that can support flexible learning opportunities. Despite their benefits, LMSs fail to cater for individual learning behavior and needs and support individualised prediction and progression. Learning Analytics (LAs) support these gaps by correlating existing learner data to provide meaningful predictive and prescriptive analysis. The industry and research community have already recognised the necessity of LAs to support modern learning needs. But a little effort has been directed towards the integration of LA into LMSs. This paper presents a novel automated Intelligence Learner Management System (iLMS) that integrates learner management and learning analytics into a single platform. The presented iLMS considers Machine Learning techniques to support learning analytics including descriptive, predictive and perspective analytics.

This paper deals with the issue of evaluating and analyzing geometric point sets in three-dimensional space. Point sets or point clouds are often the product of 3D scanners and depth sensors, which are used in the field of autonomous movement for robots and vehicles. Therefore, for the classification of point sets within an active motion, not fully generated point clouds can be used, but knowledge can be extracted from the raw impulses of the respective time points. Attractors consisting of a continuum of stationary states and hysteretic memories can be used to couple multiple inputs over time given non-independent output quantities of a classifier and applied to suitable neural networks. In this paper, we show a way to assign input point clouds to sets of classes using hysteretic memories, which are transferable to neural networks.

Cyber Supply Chain (CSC) security requires a secure integrated network among the sub-systems of the inbound and outbound chains. Adversaries are deploying various penetration and manipulation attacks on an CSC integrated network’s node. The different levels of integrations and inherent system complexities pose potential vulnerabilities and attacks that may cascade to other parts of the supply chain system. Thus, it has become imperative to implement systematic threats analyses and predication within the CSC domain to improve the overall security posture. This paper presents a unique approach that advances the current state of the art on CSC threat analysis and prediction by combining work from three areas: Cyber Threat Intelligence (CTI), Ontologies, and Machine Learning (ML). The outcome of our work shows that the conceptualization of cybersecurity using ontological theory provides clear mechanisms for understanding the correlation between the CSC security domain and enables the mapping of the ML prediction with 80% accuracy of potential cyberattacks and possible countermeasures.

Hybrid systems are dynamical systems of both continuous and discrete nature and constitute an important field of control systems theory and engineering. On the other hand, intelligent data processing has become one of the most critical devices of modern computer based systems as these systems operate in environments featuring increasing uncertainty and unpredictability. While these two approaches set completely different objectives, modern cyber-physical systems, taken as variants of hybrid systems, seem to constitute a field of increasing interest for applying intelligent techniques. Moreover, the examples of, not so recent, intelligent control systems are suggestive for considering a study on getting intelligent techniques close to hybrid systems. In this paper we present the experimental investigation we undertook in this direction. More specifically, we present and discuss the experiments carried out using Acumen a hybrid systems modeling and simulation environment. Without urging towards setting and solving questions of conceptual order we tried to figure out whether it is possible to represent intelligent behavior using a tool for modeling dynamical systems focusing on the study of its ability to permit the representation of both continuous and discrete intelligent techniques, namely, Reinforcement Learning and Hopfield neural networks. The results obtained are indicative of the problems related to the specific computational context and are useful in deriving conclusions concerning the functionality that needs to be provided by such modeling and simulation environments, in order to allow for the coexistence of hybrid systems and intelligent techniques.

This paper presents the development of regression and classification algorithms to predict greenhouse gas emissions caused by the building sector, and identify key building characteristics, which lead to excessive emissions. More specifically, two problems are addressed: the prediction of metric tons of CO2 emitted annually by a building, and building compliance to environmental laws according to its physical characteristics, such as energy, fuel, and water consumption. The experimental results show that energy use intensity and natural gas use are significant factors for decarbonizing the building sector.

Pose estimation is a key problem in computer vision, which is commonly used in augmented reality, robotics and navigation. The classical orthogonal iterative (OI) pose estimation algorithm builds its cost function based on the minimum mean square error (MMSE), which performs well when data disturbed by Gaussian noise. But even a small number of outliers will make OI unstable. In order to deal with outliers problem, in this paper, we establish a new cost function based on maximum correntropy criterion (MCC) and propose an accurate and robust correntropy-based OI (COI) pose estimation method. The proposed COI utilizes the advantages of correntropy to eliminate the bad effects of outliers, which can enhance the performance in the pose estimation problems with noise and outliers. In addition, our method does not need an extra outliers detection stage. Finally, we verify the effectiveness of our method in synthetic and real data experiments. Experimental results show that the COI can effectively combat outliers and achieve better performance than state-of-the-art algorithms, especially in the environments with a small number of outliers.

Generative Adversarial Networks (GANs) have been proposed as a method to generate multiple replicas from an original version combining a Discriminator and a Generator. The main applications of GANs have been the casual generation of audio and video content. GANs, as a neural method that generates populations of individuals, have emulated genetic algorithms based on biologically inspired operators such as mutation, crossover and selection. This paper presents the Generative Adversarial Random Neural Network (RNN) with the same features and functionality as a GAN: an RNN Generator produces individuals mapped from a latent space while the RNN Discriminator evaluates them based on the true data distribution. The Generative Adversarial RNN has been evaluated against several input vectors with different dimensions. The presented results are successful: the learning objective of the RNN Generator creates replicas at low error whereas the RNN Discriminator learning target identifies unfit individuals.

Nowadays, businesses are highly competitive as most markets are extremely saturated. As a result, customer management is of critical importance to avoid dissatisfaction that leads to customer loss. Thus, predicting customer loss is crucial to efficiently target potential churners and attempt to retain them. By classifying customers as churners and non-churners, customer loss is equated to a binary classification problem. In this paper, a new real-world dataset is used, originating from a popular web-based drug information platform, in order to predict subscriber churn. A number of methods that belong to different machine learning categories (linear, nonlinear, ensemble, neural networks) are constructed, optimized and trained on the subscription data and the results are presented and compared. This study provides a guide for solving churn prediction problems as well as a comparison of various models within the churn prediction context. The findings co-align with the notion that ensemble methods are, in principle, superior whilst every model maintains satisfying results.

Over the past few years, machine learning has played an increasingly vital role in every aspect of our society. There are countless applications of machine learning, from tradition topic such as image recognition or spam detection, to advanced areas like automatic customer service or secure automobile systems. This paper analyzes a popular machine learning application, namely housing price prediction, by applying a full machine learning process: feature extraction, data preparation, model selection, model training and optimization, and last, but not least, prediction and evaluation. We experiment with different algorithms: linear regression, random forest, and gradient boosting. This paper demonstrates the comparison of effectiveness of these algorithms that may help sellers and buyers to have a fair deal of their respective businesses.

The current study focuses on the problem of continuously tracking a dynamically evolving $$CH_4$$ C H 4 plume utilizing a mutually built consensus by heterogeneous sensing platforms: mobile and static sensors. Identifying the major complexities and emergent dynamics (leakage source, intensity, time) of such problem, a distributed, multi-agent, optimization algorithm was developed and evaluated in an indoor continuous plume-tracking application (where reaction time is critical due to the limited volume available for air saturation by the $$CH_4$$ C H 4 dispersion). The high-fidelity ANSYS Fluent suite realistic simulation environment was used to acquire the gas diffusion evolution through time. The analysis of the simulation results indicated that the proposed algorithm was capable of continuously readapting the mobile sensing platforms formation according to the density and the dispersed volume plume; combining additive information from the static sensors. Moreover, a scalability analysis with respect to the number of mobile platforms revealed the flexibility of the proposed algorithm to different numbers of available assets.

In recent years, the European Steel Industry, in particular flat steel production, is facing an increasingly competitive market situation. The product price is determined by competition, and the only way to increase profit is to reduce production and commercial costs. One method to increase production yield is to create proper scheduling for the components on the available machines, so that an order is timely completed, optimizing resource exploitation and minimizing delays. The optimization of production using efficient scheduling strategies has received ever increasing attention over time and is one of the most investigated optimization problems. The paper presents three approaches for improving flexibility of production scheduling in flat steel facilities. Each method has different scopes and modelling aspects: an auction-based multi-agent system is used to deal with production uncertainties, a multi-objective mixed-integer linear programming-based approach is applied for global optimal scheduling of resources under steady conditions, and a continuous flow model approach provides long-term production scheduling. Simulation results show the goodness of each method and their suitability to different production conditions, by highlighting their advantages and limitations.

Keyphrase extraction is a fundamental task in information management, which is often used as a preliminary step in various information retrieval and natural language processing tasks. The main contribution of this paper lies in providing a comparative assessment of prominent multilingual unsupervised keyphrase extraction methods that build on statistical (RAKE, YAKE), graph-based (TextRank, SingleRank) and deep learning (KeyBERT) methods. For the experimentations reported in this paper, we employ well-known datasets designed for keyphrase extraction from five different natural languages (English, French, Spanish, Portuguese and Polish). We use the F1 score and a partial match evaluation framework, aiming to investigate whether the number of terms of the documents and the language of each dataset affect the accuracy of the selected methods. Our experimental results reveal a set of insights about the suitability of the selected methods in texts of different sizes, as well as the performance of these methods in datasets of different languages.

Easy propagation and access to information on the web has the potential to become a serious issue when it comes to disinformation. The term “fake news” describes the intentional propagation of news with the intention to mislead and harm the public and has gained more attention recently. This paper proposes a style-based Machine Learning (ML) approach, which relies on the textual information from news, such as manually extracted lexical features e.g. part of speech counts, and evaluates the performance of several ML algorithms. We identified a subset of the best performing linguistic features, using information-based metrics, which tend to agree with the literature. We also, combined Named Entity Recognition (NER) functionality with the Frequent Pattern (FP) Growth association rule algorithm to gain a deeper perspective of the named entities used in the two classes. Both methods reinforce the claim that fake and real news have limited differences in content, setting limitations to style-based methods. Results showed that convolutional neural networks resulted in the best accuracy, outperforming the rest of the algorithms.

Text mining comprises different techniques capable to perform text analysis, information retrieval and extraction, categorization and visualization, is experiencing an increase of interest. Among these techniques, topic modeling algorithms, capable of discovering topics from large documents corpora, has many applications. In particular, considering customer experience analysis, having access to topic coherent set of opinions expressed in terms of text reviews, has an important role in both customers side and business providers. Traditional topic modeling algorithms are probabilistic models words co-occurrences oriented which can mislead topics discovery in case of short-text and context-base reviews. In this paper, we propose a customer experience analysis framework which enrich a state-of-art topic modeling algorithm (LDA) with a semantic-base topic-tuning approach.

Natural Language Processing (NLP) has evolved significantly over the last decade. This paper highlights the most important milestones of this period, while trying to pinpoint the contribution of each individual model and algorithm to the overall progress. Furthermore, it focuses on issues still remaining to be solved, emphasizing on the groundbreaking proposals of Transformers, BERT, and all the similar attention-based models.

One of the key elements in several application domains, such as policy making, addresses the scope of achieving and dealing with the very different formats, models and languages of data. The amount of data to be processed and analyzed in modern governments, organizations and businesses is staggering, thus Big Data analysis is the mean that helps organizations to harness their data and to identify new opportunities. Big Data are characterized by divergent data coming from various and heterogeneous sources and in different types, formats, and timeframes. Data interoperability addresses the ability of modern systems and mechanisms that create, exchange and consume data to have clear, shared expectations for the context, information and value of these divergent data. To this end, interoperability appears as the mean for accomplishing the interlinking of information, systems, applications and ways of working with the wealth of data. To address this challenge, in this paper a generalized and novel Enhanced Semantic Interoperability approach is proposed, the SemAI. This approach primarily focuses on the phases of the translation, the processing, the annotation, the mapping, as well as the transformation of the collected data, which have major impact on the successful aggregation, analysis, and exploitation of data across the whole policy making lifecycle. The presented prototype and its required subcomponents associated with this approach provide an example of the proposed hybrid and holistic mechanism, verifying its possible extensive application and adoption in various policy making scenarios.

Multi-stakeholder recommender systems (RSs) are a major paradigm shift from current RSs because recommendations affect not only item consumers (end-users) but also item providers (owners). They also motivate the need for new performance metrics beyond recommendation quality that explicitly affect the latter. In this work, we introduce a framework for optimizing multi-stakeholder RSs under constraints on diversity and coverage. Our goal is to make recommendations to end-users while treating each item provider equally, by ensuring sufficient user base coverage and diverse profiles of users to which items are recommended. Namely, items of each provider should be recommended to a certain number of users that are also diverse enough in their preferences. The optimization objective is that the total average rating of recommended items is as close as possible to that of a baseline RS. The problem is formulated as a quadratically constrained integer program, which is NP-Hard and impractical to solve in the presence of big data and many providers. Interestingly, we show that when only the coverage constraint exists, an instance of the problem can be solved optimally in polynomial time through its Linear Programming relaxation, and this solution can be used to initialize a low-complexity heuristic algorithm. Data experiments show good performance and demonstrate the impact of these constraints on average rating of recommended items.

Database deployment is a complex task depending on a multitude of operational parameters such as anticipated data scaling trends, expected type and volume of queries, uptime requirements, replication policies, available budget, and personnel training and experience. Thus, enterprise database administrators eventually rely on various performance metrics in conjunction to existing company policies in order to determine the best possible solution under these constraints. The recent advent of NoSQL databases, including graph databases such as Neo4j and document stores like MongoDB, added another degree of freedom in database selection since for a number of years relational databases such as PostgreSQL were the only available technology. In this work the scaling characteristics of a representative set of social queries executed on virtual machine installations of PostgreSQL and MongoDB are evaluated on a large volume of political tweets regarding Brexit. Moreover, Wiener filters for predicting the execution time of social query windows of fixed length over both databases are designed.

In the clothing industry, design, development, and procurement teams have been affected more than any other industry and are constantly under pressure to present more products with fewer resources in a shorter time. The diversity of garment designs created as new products is not found in any other industry and is almost independent of the size of the business. Science4Fashion is a semi-autonomous intelligent personal assistant for fashion product designers. Our system consists of an interactive environment where a user utilizes different modules responsible for a) data collection from online sources, b) knowledge extraction, c) clustering, and d) trend/product recommendation. This paper is focusing on two core modules of the implemented system. The Clustering Module combines various clustering algorithms and offers a consensus that arranges data in clusters. At the same time, the Product Recommender and Feedback module receives the designer’s input on different fashion products and recommends more relevant items based on their preferences. The experimental results highlight the usefulness and the efficiency of the proposed subsystems in aiding the creative fashion process.

Sentiment analysis involving the identification of sentiment polarities from textual data is a very popular area of research. Many research works that have explored and extracted sentiments from textual data such as financial news have been able to do so by employing Bidirectional Encoder Representations from Transformers (BERT) based algorithms in applications with high computational needs, and also by manually labelling sample data with help from financial experts. We propose an approach which makes possible the development of quality Natural Language Processing (NLP) models without the need for high computing power, or for inputs from financial experts on labelling focused dataset for NLP model development. Our approach introduces a two-step optimised BERT-based NLP model for extracting sentiments from financial news. Our work shows that with little effort that involves manually labelling a small but relevant and focused sample data of financial news, one could achieve a high performing and accurate multi-class NLP model on financial news.

Even though Deep Learning (DL) models are increasingly used in recent years to develop trading agents, most of them solely rely on a restricted set of input information, e.g., price time-series. However, this is in contrast with the information that is usually available to human traders that, apart from relying on price information, also take into account their prior knowledge, sentiment that is expressed regarding various markets and assets, as well as general news and forecasts. In this paper, we examine whether the use of sentiment information, as extracted by various online sources, including news articles, is beneficial when training DL agents for trading. More specifically, we provide an extensive evaluation that includes several different configurations and models, ranging from Multi-layer Perceptrons (MLPs) to Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), examining the impact of using sentiment information when developing DL models for trading applications. Apart from demonstrating that sentiment can indeed lead to improved trading efficiency, we also provide further insight on the use of sentiment-enriched data sources for cryptocurriences, such as Bitcoin, where its seems that sentiment information might actually be a stronger predictor compared to the information provided by the actual price time-series.

The complete transformation of the supply chain in a truly integrated and fully automated process, presupposes the continuous and endless collection of digital information from every stage of the production scale. The aim is not only to investigate the current situation, but also the history for every stage of the chain. Given the heterogeneity of the systems involved in the supply chain and the non-institutional interoperability in terms of hardware and software, serious objections arise as to how these systems are digitally secured. An important issue is to ensure privacy and business confidentiality. This paper presents a specialized and technologically up-to-date framework for the protection of digital security, privacy and industrial-business secrecy. At its core is Federated Learning technology, which operates over Blockchain and applies advanced encryption techniques.

This paper presents a Validation and Verification (V&V) model of Data Marketplaces. Data is extracted from the sensors embedded within the Smart city, infrastructure, or building via Application Programming Interfaces (APIs) and inserted into a Data Marketplace. The technology is based on smart contracts deployed on a private ethereum blockchain. Current issues with data in Smart cities, infrastructure, buildings, or any real estate, are the difficulty of its access and retrieval, therefore integration; its quality in terms of meaningful information; its large quantity with a reduced coverage in terms of systems and finally its authenticity, as data can be manipulated for economic advantage. In order to address these issues, this paper proposes a Data Marketplace model with a hierarchical process for data validation and verification where each stage adds a layer of data abstraction, value-added services and authenticity based on Artificial Intelligence. By using a blockchain, this presented approach is based on a decentralised method where each stakeholder stores the data. The proposed model is validated in a real application with live data: Newcastle urban observatory smart city project.