Machine Learning and Knowledge Extraction

Human activity recognition (HAR) remains an essential field of research with increasing real-world applications ranging from healthcare to industrial environments. As the volume of publications in this domain continues to grow, staying abreast of the most pertinent and innovative methodologies can be challenging. This survey provides a comprehensive overview of the state-of-the-art methods employed in HAR, embracing both classical machine learning techniques and their recent advancements. We investigate a plethora of approaches that leverage diverse input modalities including, but not limited to, accelerometer data, video sequences, and audio signals. Recognizing the challenge of navigating the vast and ever-growing HAR literature, we introduce a novel methodology that employs large language models to efficiently filter and pinpoint relevant academic papers. This not only reduces manual effort but also ensures the inclusion of the most influential works. We also provide a taxonomy of the examined literature to enable scholars to have rapid and organized access when studying HAR approaches. Through this survey, we aim to inform researchers and practitioners with a holistic understanding of the current HAR landscape, its evolution, and the promising avenues for future exploration. Full article

Dataset imbalances pose a significant challenge to predictive modeling in both medical and financial domains, where conventional strategies, including resampling and algorithmic modifications, often fail to adequately address minority class underrepresentation. This study theoretically and practically investigates how the inherent nature of medical data affects the classification of minority classes. It employs ten machine and deep learning classifiers, ranging from ensemble learners to cost-sensitive algorithms, across comparably sized medical and financial datasets. Despite these efforts, none of the classifiers achieved effective classification of the minority class in the medical dataset, with sensitivity below 5.0% and area under the curve (AUC) below 57.0%. In contrast, the similar classifiers applied to the financial dataset demonstrated strong discriminative power, with overall accuracy exceeding 95.0%, sensitivity over 73.0%, and AUC above 96.0%. This disparity underscores the unpredictable variability inherent in the nature of medical data, as exemplified by the dispersed and homogeneous distribution of the minority class among other classes in principal component analysis (PCA) graphs. The application of the synthetic minority oversampling technique (SMOTE) introduced 62 synthetic patients based on merely 20 original cases, casting doubt on its clinical validity and the representation of real-world patient variability. Furthermore, post-SMOTE feature importance analysis, utilizing SHapley Additive exPlanations (SHAP) and tree-based methods, contradicted established cerebral stroke parameters, further questioning the clinical coherence of synthetic dataset augmentation. These findings call into question the clinical validity of the SMOTE technique and underscore the urgent need for advanced modeling techniques and algorithmic innovations for predicting minority-class outcomes in medical datasets without depending on resampling strategies. This approach underscores the importance of developing methods that are not only theoretically robust but also clinically relevant and applicable to real-world clinical scenarios. Consequently, this study underscores the importance of future research efforts to bridge the gap between theoretical advancements and the practical, clinical applications of models like SMOTE in healthcare. Full article

Background. The importance of explainable artificial intelligence and machine learning (XAI/XML) is increasingly being recognized, aiming to understand how information contributes to decisions, the method’s bias, or sensitivity to data pathologies. Efforts are often directed to post hoc explanations of black box models. These approaches add additional sources for errors without resolving their shortcomings. Less effort is directed into the design of intrinsically interpretable approaches. Methods. We introduce an intrinsically interpretable methodology motivated by ensemble learning: the League of Experts (LoE) model. We establish the theoretical framework first and then deduce a modular meta algorithm. In our description, we focus primarily on classification problems. However, LoE applies equally to regression problems. Specific to classification problems, we employ classical decision trees as classifier ensembles as a particular instance. This choice facilitates the derivation of human-understandable decision rules for the underlying classification problem, which results in a derived rule learning system denoted as RuleLoE. Results. In addition to 12 KEEL classification datasets, we employ two standard datasets from particularly relevant domains—medicine and finance—to illustrate the LoE algorithm. The performance of LoE with respect to its accuracy and rule coverage is comparable to common state-of-the-art classification methods. Moreover, LoE delivers a clearly understandable set of decision rules with adjustable complexity, describing the classification problem. Conclusions. LoE is a reliable method for classification and regression problems with an accuracy that seems to be appropriate for situations in which underlying causalities are in the center of interest rather than just accurate predictions or classifications. Full article

Effective data reduction must retain the greatest possible amount of informative content of the data under examination. Feature selection is the default for dimensionality reduction, as the relevant features of a dataset are usually retained through this method. In this study, we used unsupervised learning to discover the top-k discriminative features present in the large multivariate IoT dataset used. We used the statistics of principal component analysis to filter the relevant features based on the ranks of the features along the principal directions while also considering the coefficients of the components. The selected number of principal components was used to decide the number of features to be selected in the SVD process. A number of experiments were conducted using different benchmark datasets, and the effectiveness of the proposed method was evaluated based on the reconstruction error. The potency of the results was verified by subjecting the algorithm to a large IoT dataset, and we compared the performance based on accuracy and reconstruction error to the results of the benchmark datasets. The performance evaluation showed consistency with the results obtained with the benchmark datasets, which were of high accuracy and low reconstruction error. Full article

An accurate prediction of fetal birth weight is crucial in ensuring safe delivery without health complications for the mother and baby. The uncertainty surrounding the fetus’s birth situation, including its weight range, can lead to significant risks for both mother and baby. As there is a standard birth weight range, if the fetus exceeds or falls below this range, it can result in considerable health problems. Although ultrasound imaging is commonly used to predict fetal weight, it does not always provide accurate readings, which may lead to unnecessary decisions such as early delivery and cesarian section. Besides that, no supporting system is available to predict the weight range in Saudi Arabia. Therefore, leveraging the available technologies to build a system that can serve as a second opinion for doctors and health professionals is essential. Machine learning (ML) offers significant advantages to numerous fields and can address various issues. As such, this study aims to utilize ML techniques to build a predictive model to predict the birthweight range of infants into low, normal, or high. For this purpose, two datasets were used: one from King Fahd University Hospital (KFHU), Saudi Arabia, and another publicly available dataset from the Institute of Electrical and Electronics Engineers (IEEE) data port. KFUH’s best result was obtained with the Extra Trees model, achieving an accuracy, precision, recall, and F1-score of 98%, with a specificity of 99%. On the other hand, using the Random Forest model, the IEEE dataset attained an accuracy, precision, recall, and F1-score of 96%, respectively, with a specificity of 98%. These results suggest that the proposed ML system can provide reliable predictions, which could be of significant value for doctors and health professionals in Saudi Arabia. Full article

Soil sampling constitutes a fundamental process in agriculture, enabling precise soil analysis and optimal fertilization. The automated selection of accurate soil sampling locations representative of a given field is critical for informed soil treatment decisions. This study leverages recent advancements in deep learning to develop efficient tools for generating soil sampling maps. We proposed two models, namely UDL and UFN, which are the results of innovations in machine learning architecture design and integration. The models are meticulously trained on a comprehensive soil sampling dataset collected from local farms in South Dakota. The data include five key attributes: aspect, flow accumulation, slope, normalized difference vegetation index, and yield. The inputs to the models consist of multispectral images, and the ground truths are highly unbalanced binary images. To address this challenge, we innovate a feature extraction technique to find patterns and characteristics from the data before using these refined features for further processing and generating soil sampling maps. Our approach is centered around building a refiner that extracts fine features and a selector that utilizes these features to produce prediction maps containing the selected optimal soil sampling locations. Our experimental results demonstrate the superiority of our tools compared to existing methods. During testing, our proposed models exhibit outstanding performance, achieving the highest mean Intersection over Union of 60.82% and mean Dice Coefficient of 73.74%. The research not only introduces an innovative tool for soil sampling but also lays the foundation for the integration of traditional and modern soil sampling methods. This work provides a promising solution for precision agriculture and soil management. Full article

The literature on ECG delineation algorithms has seen significant growth in recent decades. However, several challenges still need to be addressed. This work aims to propose a lightweight R-peak-detection algorithm that does not require pre-setting and performs classification on a sample-by-sample basis. The novelty of the proposed approach lies in the utilization of the typicality eccentricity detection anomaly (TEDA) algorithm for R-peak detection. The proposed method for R-peak detection consists of three phases. Firstly, the ECG signal is preprocessed by calculating the signal’s slope and applying filtering techniques. Next, the preprocessed signal is inputted into the TEDA algorithm for R-peak estimation. Finally, in the third and last step, the R-peak identification is carried out. To evaluate the effectiveness of the proposed technique, experiments were conducted on the MIT-BIH arrhythmia database (MIT-AD) for R-peak detection and validation. The results of the study demonstrated that the proposed evolutive algorithm achieved a sensitivity (Se in %), positive predictivity (+P in %), and accuracy (ACC in %) of 95.45%, 99.61%, and 95.09%, respectively, with a tolerance (TOL) of 100 milliseconds. One key advantage of the proposed technique is its low computational complexity, as it is based on a statistical framework calculated recursively. It employs the concepts of typicity and eccentricity to determine whether a given sample is normal or abnormal within the dataset. Unlike most traditional methods, it does not require signal buffering or windowing. Furthermore, the proposed technique employs simple decision rules rather than heuristic approaches, further contributing to its computational efficiency. Full article

In this review, we compiled convolutional neural network (CNN) methods which have the potential to automate the manual, costly and error-prone processing of medical images. We attempted to provide a thorough survey of improved architectures, popular frameworks, activation functions, ensemble techniques, hyperparameter optimizations, performance metrics, relevant datasets and data preprocessing strategies that can be used to design robust CNN models. We also used machine learning algorithms for the statistical modeling of the current literature to uncover latent topics, method gaps, prevalent themes and potential future advancements. The statistical modeling results indicate a temporal shift in favor of improved CNN designs, such as a shift from the use of a CNN architecture to a CNN-transformer hybrid. The insights from statistical modeling point that the surge of CNN practitioners into the medical imaging field, partly driven by the COVID-19 challenge, catalyzed the use of CNN methods for detecting and diagnosing pathological conditions. This phenomenon likely contributed to the sharp increase in the number of publications on the use of CNNs for medical imaging, both during and after the pandemic. Overall, the existing literature has certain gaps in scope with respect to the design and optimization of CNN architectures and methods specifically for medical imaging. Additionally, there is a lack of post hoc explainability of CNN models and slow progress in adopting CNNs for low-resource medical imaging. This review ends with a list of open research questions that have been identified through statistical modeling and recommendations that can potentially help set up more robust, improved and reproducible CNN experiments for medical imaging. Full article

AIM: In this study, we use Artificial Intelligence (AI), including Machine (ML) and Deep Learning (DL), to predict the long-term survival of resectable esophageal cancer (EC) patients in a high-volume surgical center. Our objective is to evaluate the predictive efficacy of AI methods for survival prognosis across different time points of oncological treatment. This involves comparing models trained with clinical data, integrating either Tumor, Node, Metastasis (TNM) classification or tumor biomarker analysis, for long-term survival predictions. METHODS: In this retrospective study, 1002 patients diagnosed with EC between 1996 and 2021 were analyzed. The original dataset comprised 55 pre- and postoperative patient characteristics and 55 immunohistochemically evaluated biomarkers following surgical intervention. To predict the five-year survival status, four AI methods (Random Forest RF, XG Boost XG, Artificial Neural Network ANN, TabNet TN) and Logistic Regression (LR) were employed. The models were trained using three predefined subsets of the training dataset as follows: (I) the baseline dataset (BL) consisting of pre-, intra-, and postoperative data, including the TNM but excluding tumor biomarkers, (II) clinical data accessible at the time of the initial diagnostic workup (primary staging dataset, PS), and (III) the PS dataset including tumor biomarkers from tissue microarrays (PS + biomarkers), excluding TNM status. We used permutation feature importance for feature selection to identify only important variables for AI-driven reduced datasets and subsequent model retraining. RESULTS: Model training on the BL dataset demonstrated similar predictive performances for all models (Accuracy, ACC: 0.73/0.74/0.76/0.75/0.73; AUC: 0.78/0.82/0.83/0.80/0.79 RF/XG/ANN/TN/LR, respectively). The predictive performance and generalizability declined when the models were trained with the PS dataset. Surprisingly, the inclusion of biomarkers in the PS dataset for model training led to improved predictions (PS dataset vs. PS dataset + biomarkers; ACC: 0.70 vs. 0.77/0.73 vs. 0.79/0.71 vs. 0.75/0.69 vs. 0.72/0.63 vs. 0.66; AUC: 0.77 vs. 0.83/0.80 vs. 0.85/0.76 vs. 0.86/0.70 vs. 0.76/0.70 vs. 0.69 RF/XG/ANN/TN/LR, respectively). The AI models outperformed LR when trained with the PS datasets. The important features shared after AI-driven feature selection in all models trained with the BL dataset included histopathological lymph node status (pN), histopathological tumor size (pT), clinical tumor size (cT), age at the time of surgery, and postoperative tracheostomy. Following training with the PS dataset with biomarkers, the important predictive features included patient age at the time of surgery, TP-53 gene mutation, Mesothelin expression, thymidine phosphorylase (TYMP) expression, NANOG homebox protein expression, and indoleamine 2,3-dioxygenase (IDO) expressed on tumor-infiltrating lymphocytes, as well as tumor-infiltrating Mast- and Natural killer cells. CONCLUSION: Different AI methods similarly predict the long-term survival status of patients with EC and outperform LR, the state-of-the-art classification model. Survival status can be predicted with similar predictive performance with patient data at an early stage of treatment when utilizing additional biomarker analysis. This suggests that individual survival predictions can be made early in cancer treatment by utilizing biomarkers, reducing the necessity for the pathological TNM status post-surgery. This study identifies important features for survival predictions that vary depending on the timing of oncological treatment. Full article

Although machine learning models are widely used in critical domains, their complexity and poor interpretability remain problematic. Decision trees (DTs) and rule-based models are known for their interpretability, and numerous studies have investigated techniques for approximating tree ensembles using DTs or rule sets, even though these approximators often overlook interpretability. These methods generate three types of rule sets: DT based, unordered, and decision list based. However, very few metrics exist that can distinguish and compare these rule sets. Therefore, the present study proposes an interpretability metric to allow for comparisons of interpretability between different rule sets and investigates the interpretability of the rules generated by the tree ensemble approximators. We compare these rule sets with the Recursive-Rule eXtraction algorithm (Re-RX) with J48graft to offer insights into the interpretability gap. The results indicate that Re-RX with J48graft can handle categorical and numerical attributes separately, has simple rules, and achieves a high interpretability, even when the number of rules is large. RuleCOSI+, a state-of-the-art method, showed significantly lower results regarding interpretability, but had the smallest number of rules. Full article

Decades of drug development research have explored a vast chemical space for highly active compounds. The exponential growth of virtual libraries enables easy access to billions of synthesizable molecules. Computational modeling, particularly molecular docking, utilizes physics-based calculations to prioritize molecules for synthesis and testing. Nevertheless, the molecular docking process often yields docking poses with favorable scores that prove to be inaccurate with experimental testing. To address these issues, several approaches using machine learning (ML) have been proposed to filter incorrect poses based on the crystal structures. However, most of the methods are limited by the availability of structure data. Here, we propose a new pose classification approach, PECAN2 (Pose Classification with 3D Atomic Network 2), without the need for crystal structures, based on a 3D atomic neural network with Point Cloud Network (PCN). The new approach uses the correlation between docking scores and experimental data to assign labels, instead of relying on the crystal structures. We validate the proposed classifier on multiple datasets including human mu, delta, and kappa opioid receptors and SARS-CoV-2 Mpro. Our results demonstrate that leveraging the correlation between docking scores and experimental data alone enhances molecular docking performance by filtering out false positives and false negatives. Full article

Digital breast tomosynthesis (DBT) is a 3D breast cancer screening technique that can overcome the limitations of standard 2D digital mammography. However, DBT images often suffer from artifacts stemming from acquisition conditions, a limited angular range, and low radiation doses. These artifacts have the potential to degrade the performance of automated breast tumor classification tools. Notably, most existing automated breast tumor classification methods do not consider the effect of DBT image quality when designing the classification models. In contrast, this paper introduces a novel deep learning-based framework for classifying breast tumors in DBT images. This framework combines global image quality-aware features with tumor texture descriptors. The proposed approach employs a two-branch model: in the top branch, a deep convolutional neural network (CNN) model is trained to extract robust features from the region of interest that includes the tumor. In the bottom branch, a deep learning model named TomoQA is trained to extract global image quality-aware features from input DBT images. The quality-aware features and the tumor descriptors are then combined and fed into a fully-connected layer to classify breast tumors as benign or malignant. The unique advantage of this model is the combination of DBT image quality-aware features with tumor texture descriptors, which helps accurately classify breast tumors as benign or malignant. Experimental results on a publicly available DBT image dataset demonstrate that the proposed framework achieves superior breast tumor classification results, outperforming all existing deep learning-based methods. Full article

Artificial neural networks (ANNs) have proven to be among the most important artificial intelligence (AI) techniques in educational applications, providing adaptive educational services. However, their educational potential is limited in practice due to challenges such as the following: (i) the difficulties in incorporating symbolic educational knowledge (e.g., causal relationships and practitioners’ knowledge) in their development, (ii) a propensity to learn and reflect biases, and (iii) a lack of interpretability. As education is classified as a ‘high-risk’ domain under recent regulatory frameworks like the EU AI Act—highlighting its influence on individual futures and discrimination risks—integrating educational insights into ANNs is essential. This ensures that AI applications adhere to essential educational restrictions and provide interpretable predictions. This research introduces NSAI, a neural-symbolic AI approach that integrates neural networks with knowledge representation and symbolic reasoning. It injects and extracts educational knowledge into and from deep neural networks to model learners’ computational thinking, aiming to enhance personalized learning and develop computational thinking skills. Our findings revealed that the NSAI approach demonstrates better generalizability compared to deep neural networks trained on both original training data and data enriched by SMOTE and autoencoder methods. More importantly, we found that, unlike traditional deep neural networks, which mainly relied on spurious correlations in their predictions, the NSAI approach prioritizes the development of robust representations that accurately capture causal relationships between inputs and outputs. This focus significantly reduces the reinforcement of biases and prevents misleading correlations in the models. Furthermore, our research showed that the NSAI approach enables the extraction of rules from the trained network, facilitating interpretation and reasoning during the path to predictions, as well as refining the initial educational knowledge. These findings imply that neural-symbolic AI not only overcomes the limitations of ANNs in education but also holds broader potential for transforming educational practices and outcomes through trustworthy and interpretable applications. Full article

Hybrid machine learning encompasses predefinition of rules and ongoing learning from data. Human organizations can implement hybrid machine learning (HML) to automate some of their operations. Human organizations need to ensure that their HML implementations are aligned with human ethical requirements as defined in laws, regulations, standards, etc. The purpose of the study reported here was to investigate technical opportunities for representing human ethical requirements in HML. The study sought to represent two types of human ethical requirements in HML: locally simple and locally complex. The locally simple case is road traffic regulations. This can be considered to be a relatively simple case because human ethical requirements for road safety, such as stopping at red traffic lights, are defined clearly and have limited scope for personal interpretation. The locally complex case is diagnosis procedures for functional disorders, which can include medically unexplained symptoms. This case can be considered to be locally complex because human ethical requirements for functional disorder healthcare are less well defined and are more subject to personal interpretation. Representations were made in a type of HML called Algebraic Machine Learning. Our findings indicate that there are technical opportunities to represent human ethical requirements in HML because of its combination of human-defined top down rules and bottom up data-driven learning. However, our findings also indicate that there are limitations to representing human ethical requirements: irrespective of what type of machine learning is used. These limitations arise from fundamental challenges in defining complex ethical requirements, and from potential for opposing interpretations of their implementation. Furthermore, locally simple ethical requirements can contribute to wider ethical complexity. Full article

Most rule induction algorithms generate rules with simple logical conditions based on equality or inequality relations. This feature limits their ability to discover complex dependencies that may exist in data. This article presents an extension to the sequential covering rule induction algorithm that allows it to generate complex and M-of-N conditions within the premises of rules. The proposed methodology uncovers complex patterns in data that are not adequately expressed by rules with simple conditions. The novel two-phase approach efficiently generates M-of-N conditions by analysing frequent sets in previously induced simple and complex rule conditions. The presented method allows rule induction for classification, regression and survival problems. Extensive experiments on various public datasets show that the proposed method often leads to more concise rulesets compared to those using only simple conditions. Importantly, the inclusion of complex conditions and M-of-N conditions has no statistically significant negative impact on the predictive ability of the ruleset. Experimental results and a ready-to-use implementation are available in the GitHub repository. The proposed algorithm can potentially serve as a valuable tool for knowledge discovery and facilitate the interpretation of rule-based models by making them more concise. Full article

Squash is a sport where referee decisions are essential to the game. However, these decisions are very subjective in nature. Disputes, both from the players and the audience, regularly occur because the referee made a controversial call. In this study, we propose automating the referee decision process through machine learning. We trained neural networks to predict such decisions using data from 400 referee decisions acquired through extensive video footage reviewing and labeling. Six positional values were extracted, including the attacking player’s position, the retreating player’s position, the ball’s position in the frame, the ball’s projected first bounce, the ball’s projected second bounce, and the attacking player’s racket head position. We calculated nine additional distance values, such as the distance between players and the distance from the attacking player’s racket head to the ball’s path. Models were trained on Wolfram Mathematica and Python using these values. The best Wolfram Mathematica model and the best Python model achieved accuracies of 86% ± 3.03% and 85.2% ± 5.1%, respectively. These accuracies surpass 85%, demonstrating near-human performance. Our model has great potential for improvement as it is currently trained with limited, unbalanced data (400 decisions) and lacks crucial data points such as time and speed. The performance of our model is almost surely going to improve significantly with a larger training dataset. Unlike human referees, machine learning models follow a consistent standard, have unlimited attention spans, and make decisions instantly. If the accuracy is improved in the future, the model can potentially serve as an extra refereeing official for both professional and amateur squash matches. Both the analysis of referee decisions in squash and the proposal to automate the process using machine learning is unique to this study. Full article

Alzheimer’s disease (AD) is a pressing global issue, demanding effective diagnostic approaches. This systematic review surveys the recent literature (2018 onwards) to illuminate the current landscape of AD detection via deep learning. Focusing on neuroimaging, this study explores single- and multi-modality investigations, delving into biomarkers, features, and preprocessing techniques. Various deep models, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and generative models, are evaluated for their AD detection performance. Challenges such as limited datasets and training procedures persist. Emphasis is placed on the need to differentiate AD from similar brain patterns, necessitating discriminative feature representations. This review highlights deep learning’s potential and limitations in AD detection, underscoring dataset importance. Future directions involve benchmark platform development for streamlined comparisons. In conclusion, while deep learning holds promise for accurate AD detection, refining models and methods is crucial to tackle challenges and enhance diagnostic precision. Full article

Complex documents have text, figures, tables, and other elements. The classification of scanned copies of different categories of complex documents like memos, newspapers, letters, and more is essential for rapid digitization. However, this task is very challenging as most scanned complex documents look similar. This is because all documents have similar colors of the page and letters, similar textures for all papers, and very few contrasting features. Several attempts have been made in the state of the art to classify complex documents; however, only a few of these works have addressed the classification of complex documents with similar features, and among these, the performances could be more satisfactory. To overcome this, this paper presents a method to use an optical character reader to extract the texts. It proposes a multi-headed model to combine vision-based transfer learning and natural-language-based Transformers within the same network for simultaneous training for different inputs and optimizers in specific parts of the network. A subset of the Ryers Vision Lab Complex Document Information Processing dataset containing 16 different document classes was used to evaluate the performances. The proposed multi-headed VisFormers network classified the documents with up to 94.2% accuracy, while a regular natural-language-processing-based Transformer network achieved 83%, and vision-based VGG19 transfer learning could achieve only up to 90% accuracy. The model deployment can help sort the scanned copies of various documents into different categories. Full article

Perovskite materials have attracted much attention in recent years due to their high performance, especially in the field of photovoltaics. However, the dark side of these materials is their poor stability, which poses a huge challenge to their practical applications. Double perovskite compounds, on the other hand, can show more stability as a result of their specific structure. One of the key properties of both perovskite and double perovskite is their tunable band gap, which can be determined using different techniques. Density functional theory (DFT), for instance, offers the potential to intelligently direct experimental investigation activities and predict various properties, including band gap. In reality, however, it is still difficult to anticipate the energy band gap from first principles, and accurate results often require more expensive methods such as hybrid functional or GW methods. In this paper, we present our development of high-throughput supervised ensemble learning-based methods: random forest, XGBoost, and Light GBM using a database of 1306 double perovskites materials to predict the energy band gap. Based on elemental properties, characteristics have been vectorized from chemical compositions. Our findings demonstrate the efficiency of ensemble learning methods and imply that scientists would benefit from recently employed methods in materials informatics. Full article