ICT Innovations 2023. Learning: Humans, Theory, Machines, and Data

Massive volumes of finance-related data are created on the Internet daily, whether on question-answering forums, news articles, or stocks analysis sites. This data can be critical in the decision-making process for targeting investments in the stock market. Our research paper aims to extract information from such sources in order to utilize the volumes of data, which is impossible to process manually. In particular, analysts’ ratings on the stocks of well-known companies are considered data of interest. Two subdomains of Information Extraction will be performed on the analysts’ ratings, Named Entity Recognition and Relation Extraction. The former is a technique for extracting entities from a raw text, giving us insights into phrases that have a special meaning in the domain of interest. However, apart from the actual positions and labels of those phrases, it lacks the ability to explain the mutual relations between them, bringing up the necessity of the latter model, which explains the semantic relationships between entities and enriches the amount of information we can extract when stacked on top of the Named Entity Recognition model. This study is based on the employment of different models for word embedding and different Deep Learning classification architectures for extracting the entities and predicting relations between them. Furthermore, the multilingual abilities of a joint pipeline are being explored by combining English and German corpora. For both subtasks, we record state-of-the-art performances of 97.69% \(F_1\) score for named entity recognition and 89.70% \(F_1\) score for relation extraction.

Using dataset analysis as a research method is becoming more popular among many researchers with diverse data collection and analysis backgrounds. This paper provides the first publicly available dataset consisting of audio segments and appropriate textual transcription in the Macedonian language. It is appropriately preprocessed and prepared for direct utilization in the automatic speech recognition pipelines. The dataset was created by students at the Faculty of Computer Science and Engineering as part of the elective course, ‘Digital Libraries’, with the audio segments sourced from a YouTube channel.

Tumor metastasis is the major cause of cancer fatality. Taking this perspective into account, the examination of gene expressions within malignant cells and the alterations in their transcriptome hold significance in the investigation of the molecular mechanisms and cellular phenomena associated with tumor metastasis. Accurately assessing a patient’s cancer condition and predicting their prognosis constitutes the central hurdle in formulating an effective therapeutic schedule for them. In recent years, a variety of machine learning techniques have widely contributed to analyzing empirical gene expression data from actual biological contexts, predicting medical outcomes, and supporting decision-making processes. This paper focuses on extracting important genes linked with each of the most common metastasis sites for breast cancer. Furthermore, the implications of the expression levels of each of the identified sets of bio-markers on the probability of predicting the occurrence of a certain metastasis are illustrated using the Shapley values as a model’s explainability framework - an approach that has never been applied on this problem before, unveils novel insights and directions for future research. The pioneering advancements of this research lie in the application of specific feature selection methods and compatible evaluation metrics to produce a small set of bio-markers for targeting a specific metastasis site, and further performing explanatory analysis of the impact of gene expression values on each of the examined metastasis sites.

Blood pressure estimation is crucial for early detection and prevention of many cardiovascular diseases. This paper explores the potential of the relatively new transformer architecture for accomplishing this task in the domain of biological signal processing. Several preceding studies of blood pressure estimation solely for PPG signals have had success with CNN and LSTM neural networks. In this study two types of transformer variants are considered: the time series and the convolutional vision transformers. The results obtained from our research indicate that this type of approach may be unsuitable for the task. However, further research is needed to make a definitive claim, since only simple transformer type are considered.

The research hypothesis in this study is that different random number generator seeds using 1D Convolutional Neural Networks impact the performance results by more than 15% on the heartbeat classification performance. Furthermore, we address a research question to evaluate the impact level of random values in the initialization of model parameters experimenting on the classification of ventricular heartbeats in electrocardiogram training and evaluating models with various feature sets based on the width of the measured samples surrounding a heartbeat location. Specific test cases consist of differently selected initial neural network parameters guided by manually selected random number seeds while preserving the rest of the training environment and hyper-parameters. We examine the influence of the random number seed on the model’s learning dynamics and ultimate F1 score on the performance of the testing dataset and conclude fluctuations resulting in 24.61% root mean square error from the average. Furthermore, we conclude that optimizing the validation in the training process does not optimize the performance in the testing. The research results contribute a novel viewpoint to the field, paving the way for more efficient and accurate heartbeat classification systems and improving diagnostic and prognostic performance in cardiac health.

The speed of information transmission in our modern day and age requires us to stay connected more than ever. With seemingly endless forms of communication, our primary digital methods of email, telephony, and message services are not without vulnerabilities. Disaster can happen at any time, whether it be man-made or natural. Often, disaster results in the loss of electrical power, internet or telecommunication infrastructures. There currently exist several emergency communication protocols and many others in development but even they have their limitations and may not function under all circumstances. With this we propose incorporating the principles of dew computing to build a reliable, stable, and resilient mobile messaging application that increases the rate of successful transmissions. This paper explores dew computing’s concepts of Independence and web service Collaboration by developing an application-level routing protocol we call Spatial-Temporal Connection (STC). By storing and relaying messages from both the primary mobile device and any neighboring devices, STC provides an alternate approach to staying connected when client-server network infrastructure breaks down.

Dew and edge computing are the sophisticated post-cloud architectural approaches that bring computing closer to the user for applications addressing the Internet of Things. In this paper, we analyze the requirements of post-cloud architectures to build such a solution, which clarify the main differences between dew and edge computing approaches. The analysis includes energy consumption, communication and processing requirements, latency, and throughput, and the evaluation shows how these requirements impact performance. In addition, we also analyze architectural approaches, including hardware/software coexistence, scalability, hardwareless computing, virtualization, interoperability, and portability. This research will check the validity of a hypothesis whether the dew and edge computing two sides of the same modern Internet of Things solution.

Leveraging advancements in information technology and the inherent interest of children with autism in robots and technology, this study explores the crucial role of analyzing application logs in enhancing therapy experiences for children with autism. By examining these logs, valuable insights can be obtained, enabling performance tracking, evidence-based evaluation, personalization of interventions, and continuous improvement. This will allow us to get more information about children’s preferences and behavior even when we are not in direct contact with them, by extending onsite robot therapies to the home environment. This research contributes to the understanding of the transformative power of log analysis and its implications for optimizing therapy experiences and advancing treatment for children with autism.

Semantic segmentation of remote sensing images is a vital task in the field of remote sensing and computer vision. The goal is to produce a dense pixel-wise segmentation map of an image, where a specific class is assigned to each pixel, enabling detailed analysis and understanding of the Earth’s surface. This paper provides an overview of semantic segmentation in remote sensing, starting with a definition of the task and its significance in extracting valuable information from remote sensing imagery. Various methods used for semantic segmentation in remote sensing are discussed, including traditional approaches such as region-based and pixel-based methods, as well as more recent deep learning-based techniques. Next, the paper delves into the available datasets for semantic segmentation of remote sensing images. Many available datasets are reviewed, highlighting their characteristics, including the number of images, image size, number of labels, spatial resolution, format and spectral bands. These datasets serve as valuable resources for training, evaluating, and benchmarking semantic segmentation algorithms in remote sensing applications. Furthermore, the paper highlights the broad range of applications enabled by semantic segmentation in remote sensing, including urban planning, land cover mapping, disaster management, environmental monitoring, and precision agriculture. Overall, this paper serves as a comprehensive guide to semantic segmentation of remote sensing images, providing insights into its definition, methods, available datasets and wide-ranging applications.

This paper investigates the application of machine learning and deep learning models to predict knee meniscus damage from magnetic resonance imaging (MRI) scans. We utilized the MRNet dataset, and processed it with different approaches, using a one-dimensional grayscale, RGB, and segmented images, complemented with features extracted using Histogram of Oriented Gradients (HOG) and Scale-Invariant Feature Transform (SIFT) techniques. Our objective was to evaluate whether a DL model could match or exceed the diagnostic performance of clinical experts such as general radiologists and orthopedic surgeons. Our findings demonstrate that our ML and DL models can predict meniscal tears with comparable accuracy to that of general medical doctors. This suggests that ML and DL models have potential to deliver rapid preliminary results post-MRI exams and augment the quality of MRI diagnoses, particularly in settings lacking specialist radiologists. Thus, integrating ML and DL models into clinical practice could enhance the quality and consistency of MRI interpretation for knee meniscus damage.

Persistent Homology (PH), a key tool in Topological Data Analysis (TDA), has gained significant traction in Machine Learning and Data Science applications in recent years. By combining techniques from algebraic topology, statistics, and computer science, PH captures the topological characteristics of datasets. This study aims to propose new classification models that integrate deep learning and Persistent Homology, exploring the impact of PH on model performance. Additionally, a transfer learning approach incorporating pre-trained networks and topological signatures is evaluated. Real-world datasets are used to assess the effectiveness of these models. The findings contribute to understanding the role of Persistent Homology in improving classification models, bridging the gap between deep learning, topological analysis, and practical data analysis. The performance of the models that include topological signatures showed better performance than the models that do not.

Natural gas is a critical commodity in the global economy, and its trade dynamics and price movements are of significant interest, particularly during times of major economic disruptions. In this research, we investigate the natural gas trade from 2017–2022 using UN Comtrade data. Our goal is to identify patterns in countries’ reliance on specific gas exporters and their strategies for reducing the risk of supply disruptions. To achieve this, we analyze trade flows between countries using graph theory methods and construct networks that illustrate these flows. Our findings indicate that the gas trade network has become more interconnected over time, suggesting increasing globalization. In addition, we create year-over-year (YoY) networks that capture changes in natural gas prices for each year. Our analysis shows that, while natural gas prices have generally increased over time, there was a decrease in the price of gas exports from the Russian Federation to Serbia and Armenia in 2022 compared to 2021. Our research provides insights into the evolution of the natural gas trade and its price fluctuations, and the proposed methodology can be extended to other globally important commodities.

Unlike traditional development of new drugs that rely on labor- and time-intensive research and clinical trials, computational approaches, deep learning technologies, in particular, have been prominent in recent research on the topic. By utilizing the ever-growing biomedical knowledge repositories and exploiting the relationship between diverse types of information (e.g., proteins, genes, molecular, diseases, drugs), graph neural networks (GNNs) primed for processing graph-structured data have a real potential for advancing the critical endeavor of drug discovery. Safe and effective drug therapy would also rely on early identification of unwanted and potentially harmful adverse effects a certain drug has on patient’s health. Hence, two, rather contrastive tasks that pertain to the process of drug discovery have been of special interest in this research. The first one is drug repurposing and the second one, a closely-related task of identifying drugs that have an adverse or negative effect on patient health namely drug-induced diseases. In this research, the task of discovering new links between drugs and diseases has been formalized as a link prediction task in a heterogenous graph. The predictive models for drug discovery proposed in this paper were tested on the ogbl-biokg (https://​ogb.​stanford.​edu/​docs/​linkprop/​#ogbl-biokg) dataset from the collection of large benchmark dataset Open Graph Benchmark (OGB) [15]. The openness and multi-source heterogeneity of the OGB dataset has provided us with an opportunity to experiment with HinSage [28], a method for inductive representational learning in heterogenous graphs. Two models based on HinSage, have been proposed proving their superior performance when compared with more traditional similarity-based baseline methods. Furthermore, a selected newly discovered relationship with a potential for drug repurposing has been discussed through the lenses of related clinical-experimental trials.

Multiplex collaboration networks facilitate intricate connections among individuals, enabling multidimensional collaborations across various domains and fostering synergistic knowledge exchange. This study focuses on the construction and basic analysis of a multiplex collaboration network among employees at the Faculty of Computer Science and Engineering (FCSE), Ss. Cyril and Methodius University in Skopje. The multiplex network is built with three layers based on: scientific collaborations resulting from joint project participations by FCSE employees, joint employees participations in the FCSE graduation thesis committees, and scientific FCSE employees collaborations defined by co-authorships in Google Scholar papers.

The network’s structure plays a vital role in determining the information accessibility and cooperative opportunities for individuals within FCSE institution. The aim here is to investigate the FCSE multiplex collaboration network’s internal structure for discovering its latent knowledge and understand its implications. We perform identification of key individuals within the network, by computing various centrality and hubs detection network metrics. Additionally, we employ a community detection algorithm to reveal the underlying modular structure of the network.

By comprehensively analyzing the acquired multiplex collaboration network model, we contribute to a better understanding of the collaboration patterns among FCSE employees. The findings can potentially inform decision-making processes and foster strategic planning aimed at enhancing collaboration and knowledge sharing within the institution.

Social media resurgence of antisocial behavior has exerted a downward spiral on stereotypical beliefs, and hateful comments towards individuals and social groups, as well as false or distorted news. The advances in graph neural networks employed on massive quantities of graph-structured data raise high hopes for the future of mediating communication on social media platforms. An approach based on graph convolutional data was employed to better capture the dependencies between the heterogeneous types of data.

Utilizing past and present experiences on the topic, we proposed and evaluated a graph-based approach for antisocial behavior detection, with general applicability that is both language- and context-independent. In this research, we carried out an experimental validation of our graph-based approach on several PAN datasets provided as part of their shared tasks, that enable the discussion of the results obtained by the proposed solution.

Quasigroups have various applications in mathematics, computer science, and cryptography. In coding theory and cryptography they have been used in error-correcting codes, error-detection codes, to construct key exchange protocols and cryptographic primitives. There are also used in graph theory, experimental design and combinatorial designs. Quasigroups of order \(2^n\) can be represented as vector valued Boolean functions from \(\{0,1\}^n \times \{0,1\}^n\) to \(\{0,1\}^n\). When the order of each coordinate functions is at most two, they are called Multivariate Quadratic Quasigroups (MQQ). In this paper we give a description of the functions representing Bilinear MQQ quasigroups, with a special focus on quasigroups that are commutative or associative.

Motivated by the need of reducing the huge amount of data navigating simultaneously through a network of polarized splicing processors, we look to the possibility of introducing probabilities which theoretically could decrease this amount, at a price of some loss of certainty. We imagined two possible situations regarding the splicing step: to associate either fixed or dynamically computed probabilities with splicing rules in every node. Similarly to the splicing step, two situations could be considered for the communication step depending on the way the probabilities are associated: statically or dynamically. We believe that this new feature together with the communication protocol based on polarization might facilitate software simulations or hardware implementations.