Search Results (128)

The purpose of this article is to conduct a systematic literature review on the role of deep learning in credit risk prediction for fintech lending, with particular emphasis on model architectures, imbalanced data handling techniques, and the mathematical foundations underpinning these methods. Open-access scientific publications retrieved from three complementary databases (Scopus, IEEE Xplore Digital Library, and Web of Science Core Collection) were used to conduct the systematic literature review. Following the PRISMA 2020 protocol, 30 publications were selected after a rigorous multi-stage screening process involving deduplication across databases, temporal filtering (2015–2026), and thematic eligibility assessment. Data for analysis were processed using Python-based bibliometric tools and network analysis (replicating R Bibliometrix and VOSviewer functionalities). The results of the analysis indicate a sustained growth in research on deep learning applications for fintech credit risk, with a compound annual growth rate (CAGR) of approximately 29.2% and an average of 27.17 citations per document. The segmentation of the studied conceptual landscape made it possible to identify four interconnected thematic clusters: (1) peer-to-peer lending and default prediction architectures; (2) explainability and XAI-based methods for credit scoring; (3) imbalanced data and hybrid deep-learning frameworks; and (4) credit risk assessment combining deep learning and statistical approaches. The following research areas on the deep learning-based transformation of fintech credit risk prediction have been identified: (1) feedforward deep neural networks and attention-based architectures (LSTM, CNN) as dominant predictive engines; (2) hybrid deep ensemble and deep-boosting frameworks (e.g., LightGBM-Attention, GBDT-Deep FFM) as emerging high-performance paradigms; (3) specialized techniques for imbalanced data handling—including ADASYN, SMOTE, cost-sensitive learning, and balanced stratified prioritized experience replay—as critical methodological frontiers; and (4) transformer-based and XAI-integrated architectures as the emerging frontier. The originality of this article lies in its explicit focus on the mathematical and methodological challenges of deep learning-based credit risk prediction in fintech lending, providing an actionable research agenda that addresses class imbalance, uncertainty quantification, loss function design, concept drift, and regulatory compliance. The findings provide valuable insights for scholars, practitioners, and policymakers, and outline a concrete roadmap for developing more accurate, robust, and explainable credit risk models in the rapidly evolving fintech ecosystem. Full article

Predictive quality systems are central to Industry 4.0 manufacturing, yet detecting rare defects that pass established quality gates remains an open problem. This study addresses post-quality-gate failure detection in automotive brake manufacturing, where 310 faulty units (1.20%) among 25,756 production records create a naturally occurring extreme class imbalance of 1:82. Fault labels are derived from warranty reports and linked to multi-station production line measurements, while negative samples may include latent failures, motivating a recall-focused evaluation. We propose a Quality 4.0 machine learning framework that compares five resampling methods (ADASYN, SMOTE-Tomek, KMeans-SMOTE, CTGAN, and TVAE) plus a no-resampling baseline across 24 classifiers and stacking ensembles. In total, 504 configurations are tested on a held-out test set. The proposed SVM-RBF model trained on ADASYN-augmented data achieves recall of 0.871, specificity of 0.982, balanced accuracy of 0.926, and ROC-AUC of 0.952, producing only 93 false positives (FPR = 1.8%). Stacking ensembles provide alternative operating points maximizing the detection rate (93.5%) and a separate operating point with the highest discrimination capacity (ROC-AUC = 0.975). Feature importance analysis through Permutation Importance and SHAP identifies Force Increment as the leading feature under both attribution methods. Friedman and Wilcoxon tests confirm statistically significant differences among strategies. The framework offers a practical way to add predictive capability to existing quality control systems. Full article

Accurate prediction of loan defaults is essential for financial institutions seeking to minimize credit losses and maintain portfolio stability. In the vehicle financing segment of emerging markets, real-world datasets frequently exhibit extreme class imbalance ratios that far exceed those encountered in standard benchmark corpora, posing severe challenges for conventional machine learning pipelines. This study introduces a gradient boosting framework integrating Weight of Evidence (WoE) transformation, Bayesian hyperparameter optimization, and three complementary classifiers—Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost)—to predict vehicle loan default risk. The methodology is evaluated on a large-scale, fully anonymized Turkish vehicle loan dataset ($N=207,572$) with an extreme imbalance ratio of 1:1133 (183 defaults versus 207,389 non-defaults). A strict three-way data partition (60% training, 20% validation, 20% test) is adopted to ensure leakage-free model selection and unbiased performance estimation. A multi-stage experimental pipeline is developed encompassing: (i) statistical feature selection via Mann–Whitney U and chi-square tests with adaptive thresholding, (ii) a comparative analysis of seven resampling strategies including Synthetic Minority Oversampling Technique (SMOTE) variants, Adaptive Synthetic Sampling (ADASYN), and focal loss weighting, (iii) a greedy forward selection ensemble procedure for heterogeneous model fusion, and (iv) a systematic training-set size sensitivity analysis across eight majority undersampling ratios. Under the leakage-free evaluation protocol, the highest-AUC individual model (LightGBM with SMOTE-ENN) achieves an Area Under the Curve (AUC) Receiver Operating Characteristic (ROC) of 0.710 (95% bootstrap CI: 0.614–0.798), while CatBoost with cost-sensitive weighting exhibits superior operational metrics (KS $=0.389$, PR-AUC $=0.011$). The greedy ensemble procedure exhibits high selection instability with only 37 validation-set positives, providing a methodological finding on the minimum sample requirements for reliable ensemble construction under extreme scarcity. Ablation results confirm that WoE encoding contributes 3.1 percentage points to the overall AUC gain. Tree SHAP-based interpretability analysis identifies the financing-to-age ratio, WoE-encoded occupation group, and log financing amount as the primary predictive drivers, with cross-model stability confirmed via Spearman rank correlation. A decision support analysis provides precision–recall curves, a Brier score of 0.0082, reliability diagrams, and threshold-dependent performance at operationally plausible review rates. Fairness evaluation across gender and marital status subgroups demonstrates that threshold-dependent metrics such as Disparate Impact Ratio and Equalized Odds Gap are inherently compromised under extreme minority scarcity, whereas rank-based subgroup AUC analysis with bootstrap 95% confidence intervals preserves meaningful discriminative assessment. These findings provide an empirically validated framework for credit default prediction in highly imbalanced and data-scarce financial environments. Full article

Background/Objectives: This paper aims to design an explainable and accurate ML framework to support the automatic classification of Carpal Tunnel Syndrome (CTS) severity from structured patient data. Methods: For the experiment, an open-source dataset of 1037 samples was used. Following stratified partitioning, 305 samples were held out as the test set; the remaining training set (n = 732) was augmented to 1216 balanced samples via ADASYN, yielding an 80/20 train/test ratio relative to the final dataset (n = 1521). In order to solve the problem of imbalance associated with CTS cases of moderate and severe severity, the Adaptive Synthetic Sampling (ADASYN) technique was employed. The model’s predictive capacity was increased by means of feature engineering methods, such as polynomial transformations and clinically relevant interactions. Specifically, four ensemble learning models (XGBoost, Random Forest, LightGBM, and CatBoost) were optimized and ensembled with the use of a stacking approach with a base algorithm of LightGBM. The explainability of the model was ensured through SHAP and LIME analysis. Results: As a result, the stacking ensemble was able to reach a test accuracy of 91.15%, an F1-score of 91.13%, and an ROC-AUC of 0.9708. The proposed ensemble performed superiorly compared to any other individual algorithm while having stable performance across all severity categories. Conclusions: Through the explainability analysis, it was observed that such a classification model relies on important clinically relevant predictors, including cross-sectional area (CSA), duration of symptoms, pain level measured by the numeric rating scale of pain (NRS), and palmar bowing (PB). Full article

Sentiment analysis has become increasingly important across various domains, particularly in business intelligence, where it is crucial for improving the performance of companies by identifying the sentiments and emotions expressed in customer feedback on products and services. Despite its growing relevance, sentiment analysis still faces several challenges, including class imbalance in datasets, limitations in feature extraction techniques, and the selection of appropriate classification models. Effectively addressing these challenges requires the integration of robust representation methods, reliable data balancing strategies, and efficient classification frameworks. In this study, we propose a novel sentiment analysis approach that combines SBERT for contextual feature extraction, WGAN-based synthetic data generation for addressing class imbalance, and a soft voting ensemble classifier for improved prediction. The proposed approach is evaluated on five datasets, including two English datasets and three Arabic datasets, in order to assess its performance in a multilingual setting. We compare the effectiveness of the proposed model with several baseline machine learning classifiers, as well as with commonly used data balancing techniques such as the synthetic minority over-sampling technique (SMOTE) and adaptive synthetic (ADASYN). The evaluation is conducted using multiple performance metrics, including accuracy, precision, recall, F1-score, MCC, ROC–AUC and training and inference time, along with different validation strategies including fixed train–test splits and k-fold cross-validation. The experimental results demonstrate the effectiveness and stability of the proposed approach. In particular, they highlight the importance of capturing sentence-level contextual representations and generating realistic synthetic samples to address class imbalance. Full article

Internet of Medical Things (IoMT) devices are inherently vulnerable to cyberattacks, typically due to their limited processing power and memory capacity. Their widespread use in healthcare poses a significant security risk, threatening patient data privacy and the continuity of services. This study examines the effects of data imbalance correction and balancing strategies on the performance of machine and deep learning models using openly available IoMT datasets. In this context, four different balancing methods—RandomUnderSampler, SMOTE, Borderline-SMOTE, and ADASYN—were applied to three open-access IoMT datasets: ECU-IoHT, WUSTL, and CICIoMT2024. Performance analyses were conducted using five machine learning algorithms (AdaBoost, Logistic Regression, Random Forest, XGBoost, and K-Nearest Neighbor (KNN)) and two deep learning algorithms (Convolutional Neural Networks (CNN) and Deep Neural Networks (DNN)). In the highly imbalanced binary setting of the CICIoMT2024 dataset, the combination of RandomUnderSampler and SMOTE under the balanced-training/original-testing scenario produced the strongest improvement in the binary CICIoMT2024 setting, increasing the F1-Score from the unbalanced baseline to 99.87% for Random Forest and 99.86% for XGBoost across repeated runs. However, the benefit of balancing was not universal. In datasets with stronger class separability, such as ECU-IoHT, and in several multi-class settings, the effect of balancing was limited or, in some cases, inferior to the unbalanced baseline. These findings indicate that balancing is most effective under specific conditions, particularly in highly imbalanced binary tasks, and should be validated using class-sensitive metrics rather than overall performance alone. Full article

The 316L stainless steel is widely utilized as structural material in hydrogen production industry due to its excellent combination of corrosion resistance and mechanical properties. However, it remains susceptible to localized pitting corrosion in chloride-containing high-temperature environments. Especially, the main electrolysis byproduct sodium chlorate (NaClO₃) also has complicated effect on pitting corrosion. Therefore, evaluating and predicting the pitting severity grades of 316L steel in NaClO₃ and NaCl environment is essential for controlling operation risks. In recent years, machine learning (ML) methods have gained significant attention in the field of corrosion prediction; however, existing research has primarily focused on the regression prediction of continuous parameters, while studies dedicated to the classification and evaluation of pitting severity grades remain relatively limited. Furthermore, experimental datasets are commonly constrained by small sample sizes and imbalanced class distributions, which hinder the performance enhancement of classification models. Based on experimental pitting data of 316L stainless steel, this study employs ADASYN (Adaptive Synthetic Sampling) to mitigate data imbalance and develops a Feedforward Neural Network (FFNN) for pitting rate classification. The proposed model is compared and analyzed against several commonly used machine learning models. Through a comprehensive evaluation of predictive performance, the feasibility of the developed model in pitting severity grading is verified, thereby providing a novel approach for the predictive evaluation of the pitting corrosion of 316L stainless steel. Full article

Road traffic injuries represent one of the most critical public health challenges in the Gulf region. Predicting traffic accident severity is therefore a critical component of evidence-based road safety management. In this study, we develop machine learning frameworks for predicting traffic accident severity using Qatar’s national dataset (2020–2025), addressing extreme class imbalance and interpretability. A dataset of 588,023 accident records was systematically preprocessed from 1,000,500 raw reports. We compare three approaches: multi-class (four severity levels), binary (Safe vs. Severe), and cascaded two-stage (combining both). Six classifiers were evaluated across two encoding methods and three balancing strategies. Systematic hyperparameter tuning with 5-fold stratified cross-validation was performed for all models. The binary LightGBM classifier achieved BA = 71.04%, AUC-ROC = 0.772, Sensitivity = 61.03%, and Specificity = 81.05%, demonstrating superior performance over multi-class approaches. Temporal validation on 2025 data (trained on 2020–2024 data) supported good temporal generalization. Analysis of 10,000 test instances identified the time period as the dominant predictor of accident severity. The binary LightGBM framework provides an interpretable and effective approach for severe accident identification and risk prioritization, with SHAP findings supporting targeted temporal enforcement and pedestrian safety as evidence-based policy priorities. Full article

Sentiment analysis of online game reviews has increasingly become important in understanding player experiences and supporting data-driven game development. However, research in this domain has continuously faced two unresolved challenges: (1) the extreme imbalance between positive and negative feedback, and (2) the inefficiency of existing feature-weighting schemes in capturing sentiment signals embedded in informal gaming discourses. Prior works demonstrated that negative feedback—though a few in number are highly influential—usually contain richer emotional content and longer textual structures; yet, prevailing classification models often perform poorly for these minorities (i.e., negative feedback). Numerous studies explored multimodal imbalance issues, class imbalance in cross-lingual ABSA (Aspect-Based Sentiment Analysis), reinforcement-learning-based architectures for imbalanced extraction tasks, and oversampling strategies like SMOTE (Synthetic Minority Over-sampling Technique) variants. Few investigations specifically addressed imbalanced sentiment classification in the contexts of online game reviews, where user-generated content exhibits unique lexical, structural, and emotional characteristics. To address these gaps, this study integrated TF-IDF (Term Frequency-Inverse Document Frequency), VADER (Valence Aware Dictionary and Sentiment Reasoner) lexicon features, and IGM (Inverse Gravity Moment) weightings with advanced oversampling methods such as ADASYN (Adaptive Synthetic Sampling Approach for Imbalanced Learning) and Borderline-SMOTE to improve the detection of minority sentiment classes. Ensemble models, including XGBoost (Extreme Gradient Boosting) and LightGBM (Light Gradient-Boosting Machine), were further employed to enhance the robustness of imbalance. Using a large-scale dataset of Steam game reviews, the proposed framework demonstrated substantial improvement in identifying negative sentiments, addressing a critical limitation in the existing computational game-analysis literature, and advancing the modeling for detecting the emotion-rich but imbalance-prone user feedback. Full article

Sleep plays a fundamental role in human health and cognitive functioning, motivating the development of reliable and scalable methodologies for sleep stage classification (SSC). Recent advances in non-invasive and economically sustainable sensing technologies enable continuous sleep monitoring beyond laboratory settings. However, SSC remains a challenging data analytics task due to the intrinsic class imbalance among sleep stages. This study investigates the effectiveness of different imbalanced data management strategies within a machine learning framework for non-invasive SSC. The proposed approach relies exclusively on heart rate and motion signals, which can be acquired through wearable devices or contactless under-mattress sensors, making it suitable for longitudinal monitoring scenarios. Using the PhysioNet DREAMT dataset, 32 experimental scenarios are defined by combining data-level techniques (ADASYN oversampling with different balancing weights), algorithm-level strategies (cost-sensitive learning), and hybrid solutions. Four model families are evaluated—Decision Tree, k-Nearest Neighbors, Ensemble Classifiers, and Artificial Neural Networks—across classification tasks involving 2, 3, 4, and 5 sleep stages. The experimental results show that ensemble-based models provide robust and consistent performance under severe class imbalance, achieving macro accuracies of 82% for sleep–wake detection, 73% for 3-stage classification, 72% for 4-stage classification, and 64% for 5-stage classification. These findings confirm the relevance of imbalance-aware analytics and demonstrate the feasibility of accurate, minimally invasive SSC within big data and cognitive computing paradigms. Full article

Healthcare fraud poses significant risks to insurance systems, undermining both financial sustainability and equitable access to care. Accurate detection of fraudulent claims is therefore critical to ensuring the integrity of healthcare insurance operations. However, the increasing sophistication of fraud techniques and limited data availability have undermined the performance of traditional detection approaches. To address these challenges, this paper proposes WCGAN-GA-RF, an integrated fraud detection framework that synergistically combines Wasserstein Conditional Generative Adversarial Network with gradient penalty (WCGAN-GP) for synthetic data generation, genetic algorithm-based feature selection (GA-RF) for dimensionality reduction, and Random Forest (RF) for classification. The proposed framework was empirically validated on a real-world dataset of 16,000 healthcare insurance claims from a Chinese healthcare technology firm, characterized by a 16:1 class imbalance ratio ($5.9\%$ fraudulent samples) and 118 original features. Using a stratified $80/20$ train–test split with results averaged over five independent runs, the WCGAN-GA-RF framework achieved a precision of $96.47\pm 0.5\%$, a recall of $97.05\pm 0.4\%$, and an F1-score of $96.26\pm 0.4\%$. Notably, the GA-RF component achieved a $65\%$ feature reduction (from 80 to 28 features) while maintaining competitive detection accuracy. Comparative experiments demonstrate that the proposed approach outperforms conventional oversampling methods, including Random Oversampling (ROS), Synthetic Minority Oversampling Technique (SMOTE), and Adaptive Synthetic Sampling (ADASYN), particularly in handling high-dimensional, severely imbalanced healthcare fraud data. Full article

Precise water quality forecasting is vital for sustainable resource management and public health, especially in semi-arid environments. This study investigates the predictive capabilities of ten Machine Learning (ML) algorithms using a dataset of 308 drinking water samples collected from various districts in Şanlıurfa Province, Türkiye. We evaluated ten predictive models, including Support Vector Regressor (SVR) and Extreme Gradient Boosting (XGBoost), both integrated with dimensionality reduction and hyperparameter optimization. Nineteen physicochemical and microbiological parameters—Temperature, chlorine (Cl⁻), pH, Electrical Conductivity (EC), Total Dissolved Solids (TDS), nitrite (NO₂⁻), nitrate (NO₃⁻), ammonium (NH₄⁺), sulfate (SO₄²⁻), Free Chlorine (Cl₂), calcium (Ca²⁺), magnesium (Mg²⁺), sodium (Na⁺), potassium (K⁺), fluoride (F⁻), trihalomethanes (THMs), Escherichia coli, Enterococci, Total Coliform—were used as input features. The dataset was split into training (75%) and testing (25%) subsets, and model performance was assessed through 10-fold cross-validation and hold-out testing procedures. To improve model generalization and mitigate the effects of class imbalance, we implemented the Adaptive Synthetic Sampling (ADASYN) technique. ML algorithms were evaluated using standard regression metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and the Coefficient of Determination (R²). The LSTM model optimized using Randomized Search outperformed the SVR and XGBoost models, demonstrating the highest accuracy and generalization capability, as evidenced by the superior R² value of 0.999 following ADASYN balancing and the lowest RMSE (1.206). These findings underscore the effectiveness of the LSTM framework in modeling the complex variance of the Weighted Arithmetic Water Quality Index (WAWQI). The findings of this study are expected to support future water quality monitoring strategies, inform policy development, and contribute to sustainable water resource management in arid and semi-arid regions. Full article

In today’s digital era, cyberattacks pose a critical threat to networks of all scales, from local systems to global infrastructures. Intrusion detection systems (IDSs) are essential for identifying and mitigating such threats. However, existing machine learning-based IDS often suffer from low detection accuracy, heavy reliance on manual feature extraction, and limited coverage of attack categories. To address these limitations, we propose a modular, deployment-ready intrusion detection framework that integrates multiple heterogeneous datasets through a hybrid transformer–multilayer perceptron (Transformer–MLP) architecture. The system employs three parallel Transformer–MLP models, each specialized for a distinct dataset, whose probabilistic outputs are fused using a weighted decision-level strategy. Unlike traditional feature-level fusion, this strategy ensures module independence, eliminates the need for global retraining when adding new components, and provides seamless modular scalability. The framework accurately identifies twenty-one traffic categories, including one benign and twenty attack classes, derived from a unified mapping across multiple heterogeneous sources to ensure a consistent cross-dataset taxonomy. By combining advanced contextual representation learning with ensemble-based probabilistic fusion, the framework demonstrates high detection accuracy and practical applicability in real-world network environments. The Transformer module captures complex contextual dependencies, while the MLP performs final classification. Class imbalance is mitigated via adaptive synthetic sampling (ADASYN), synthetic minority over-sampling technique (SMOTE), edited nearest neighbor (ENN), and class weight adjustments. Empirical evaluation demonstrates the framework’s high effectiveness: for binary classification, it achieves 99.98% on CICIDS2017, 99.19% on NSL-KDD, and 99.98% on NF-BoT-IoT-v2; for two-stage multi-class classification, 99.56%, 99.55%, and 97.75%; and for one-phase multi-class classification, 99.73%, 99.07%, and 98.23%, respectively. Moreover, the framework enables real-time deployment with 4.8–6.9 ms latency, 9800–14,200 fps throughput, and 412–458 MB memory. These results outperform existing multi-dataset IDS approaches, highlighting the architectural effectiveness, robustness, and practical applicability of the proposed framework. Full article

This paper examines the use of Machine Learning (ML) approaches in maintaining academic integrity using the information provided in the Moodle system logs. The paper focuses on data set construction, handling the issue of class imbalance, and the assessment of the performance of different ML models in uncovering academic fraud. Twelve different data sets were created by using the concept of temporal windows (e.g., one-day and three-day windows) during the feature extraction stage from the Moodle system logs. The manual labeling of the data sets was done based on a predefined set of rules that outline the fraudulent activities. The issue of class imbalance was treated using eleven different resampling approaches, such as SMOTE, ADASYN, Tomek Links, and NearMiss. We evaluated six classification algorithms, thus resulting in a total of 792 experiments based on the interactions between the data sets, resampling methods, and classification algorithms. The results from the experiment show that the Random Forest and AdaBoost models performed the best in the experiment. Furthermore, we observed a trade-off between fraud detection rates and model precision based on the temporal windows and resampling methods. The shortest temporal windows and hybrid undersampling approaches resulted in the maximum recall value in this study and could identify the greatest number of at-risk students. On the other hand, the longest temporal windows and hybrid oversampling approaches with data cleaning resulted in the best results in terms of F1-Score and Cohen’s Kappa statistics. The results provide conclusive evidence that the models can identify fraud; however, they should be used as predictive models for the improvement of proctoring approaches, such as random selection for verification or seating arrangement strategies, instead of judgment models. Full article

As network environments become increasingly interconnected, ensuring robust cyber-security has become critical, particularly with the growing sophistication of modern cyber threats. Intrusion detection systems (IDSs) play a vital role in identifying and mitigating unauthorized or malicious activities; however, conventional machine learning-based IDSs often rely on handcrafted features and are limited in their ability to detect diverse attack types across disparate network domains. To address these limitations, this paper introduces a novel unified intrusion detection framework that implements “Structural Dualism” to integrate three heterogeneous benchmark datasets (CSE-CIC-IDS2018, NF-BoT-IoT-v2, and IoT-23) into a harmonized, protocol-agnostic representation. The framework employs a shared autoencoder architecture with dataset-specific projection layers to learn a unified latent manifold. This 15-dimensional space captures the underlying semantics of attack patterns (e.g., volumetric vs. signaling) across multiple domains, while dataset-specific decoders preserve reconstruction fidelity through alternating multi-domain training. To identify complex micro-signatures within this manifold, the framework utilizes a synergistic hybrid convolutional neural network–deep neural network (CNN–DNN) classifier, where the CNN extracts spatial latent patterns and the DNN performs global classification across twenty-five distinct classes. Class imbalance is addressed through resampling strategies such as adaptive synthetic sampling (ADASYN) and edited nearest neighbors (ENN). Experimental results demonstrate remarkable performance, achieving 99.76% accuracy for binary classification and 99.54% accuracy for multi-class classification on the merged dataset, with strong generalization confirmed on individual datasets. These findings indicate that the shared autoencoder-based CNN–DNN framework, through its unique feature alignment and spatial extraction capabilities, significantly strengthens intrusion detection across diverse and heterogeneous environments. Full article