Search Results (3,155)

Landslide susceptibility assessment is important for disaster prevention and sustainable land-use planning in mountainous regions. However, conventional discretization methods often overlook threshold effects in conditioning factors, and many machine learning models still have limited interpretability. This study develops an integrated framework that combines ChiMerge discretization, Weight of Evidence (WOE) transformation, and tree-based ensemble learning to map landslide susceptibility in Jiuzhaigou County, Sichuan Province, China. A landslide inventory of 164 points was compiled from field investigations and hazard records, and fourteen topographic, geological, and environmental conditioning factors were derived from multi-source spatial datasets. Continuous factors were discretized using ChiMerge, a supervised chi-square-based discretization method that identifies statistically meaningful thresholds according to the distributions of landslide and non-landslide samples. WOE values were then calculated to quantify the association between each factor class and landslide occurrence. Three WOE-based ensemble models, WOE-CatBoost, WOE-LightGBM, and WOE-RF, were constructed and compared. All models showed high predictive performance (AUC > 0.90), with WOE-CatBoost performing best (AUC = 0.9432). Its high and very high susceptibility zones covered 28.59% of the study area but contained 85.96% of observed landslides. High-risk areas were mainly concentrated in steep valleys, fractured lithological zones, erosion belts, and areas affected by engineering activities, such as road construction, slope cutting, tourism infrastructure development, and settlement expansion. The proposed framework improves prediction accuracy and interpretability and provides spatial support for landslide prevention and sustainable land-use management. Full article

Determining HIV-1 tropism provides the prognosis of HIV infection and is required before prescribing maraviroc, an entry inhibitor that blocks the interaction between the viral gp120 and the CCR5 coreceptor. However, existing prediction algorithms have been developed primarily for the globally most prevalent subtypes (B, C, and CRF01_AE) and often show reduced performance for other HIV-1 genetic variants. Sub-subtype A6 and circulating recombinant form CRF63_02A6 dominate the HIV-1 epidemic in Russia and other Former Soviet Union (FSU) countries, yet the reliability of tropism prediction for these viruses remains virtually unexplored. We phenotypically determined the tropism of 25 clinical isolates (11 R5, 1 X4, and 7 dual-tropic R5/X4) using U87.CD4.CCR5 and U87.CD4.CXCR4 cell lines and performed a comparative analysis of eight existing genotypic tools (Geno2pheno, WebPSSM, T-CUP 2.0, the Delobel/Garrido rules, and others) or their modifications on a combined dataset that included Los Alamos National Laboratory (LANL) reference sequences (subtypes A, B, C, CRF01_AE, and CRF02_AG) and our laboratory-derived isolates. Most models achieved high accuracy for globally prevalent subtypes (≈95% for B, C, and CRF01_AE) but showed markedly reduced performance for sub-subtype A6 (best accuracy among existing models, 85%) and CRF63_02A6 (best accuracy, 72%), with a poor balance between sensitivity and specificity. To address this problem, we developed HIV-V3Augur, an ensemble stacking model based on the Random Forest and Support Vector Machine (SVM) machine learning algorithms, trained on Pseudo Amino Acid Composition (PseAAC) and Relative Synonymous Codon Usage (RSCU) features with 10-fold stratified cross-validation. HIV-V3Augur achieved an accuracy of 77%, sensitivity of 79%, and specificity of 79% on sub-subtype A6, and on CRF63_02A6 it reached an accuracy of 95%, sensitivity of 87%, and specificity of 100%. Cross-validation demonstrated that HIV-V3Augur represents a balanced genotypic tropism prediction tool for understudied HIV-1 variants circulating in the FSU region. HIV-V3Augur can be used locally through a graphical user interface. Full article

The snapshot proper orthogonal decomposition (POD) method relies on synchronously sampled datasets, significantly limiting its utility for analyzing asynchronous measurements in unsteady flow studies. This paper proposes a frequency-domain proper orthogonal decomposition (FDPOD) method tailored for mode extraction and flow field reconstruction from asynchronously sampled data. The FDPOD framework integrates three key components: frequency-domain transformation to decouple phase discrepancies inherent in asynchronous sampling, power spectral density (PSD) analysis combined with segmented ensemble averaging to suppress spectral leakage errors, and eigenvalue decomposition of energy-ranked frequency components to identify dominant coherent structures. Validated through numerical simulations of a subsonic jet and experimental measurements from a low-speed mixed-flow fan, the method demonstrates exceptional performance under asynchronous conditions: cumulative energy errors are reduced to 0.3% across the first 50 modes, while flow field reconstruction achieves 99.5% accuracy. Dominant mode structures exhibit remarkable consistency with those derived from synchronous conditions, with hot-wire measurement errors remaining below 0.03% for both asynchronous and temporally shuffled datasets. These results position FDPOD as a robust and practical tool for analyzing complex unsteady flows where synchronous data acquisition proves impractical, particularly in large-scale or spatially distributed measurement systems. Full article

In bank-dominated financial systems, the accumulation of non-performing loans is a recognised source of systemic vulnerability, as correlated corporate defaults can erode bank capital, impair liquidity, and propagate stress across interconnected portfolios. Firm-level default detection thus constitutes a microprudential foundation of macroprudential stability: the reliable early identification of risky borrowers reduces both individual credit losses and the aggregate exposures that drive system-level fragility. Yet the use of structured data-mining pipelines for this task remains underexplored in Central and Eastern Europe. This study applies the CRISP-DM methodology to predict corporate loan default using data on 302 Slovak corporate borrowers, combining financial ratios from publicly available financial statements with selected company and loan-related information from internal bank records. Seven individual classifiers were developed and compared: decision trees (CART, CHAID, C5.0), logistic regression, discriminant analysis, and neural networks (MLP, RBF), together with a stacked ensemble based on their outputs. Model performance was evaluated using sensitivity, overall classification accuracy, and area under the ROC curve (AUC), with sensitivity treated as the primary criterion because of the asymmetric costs of misclassification in credit risk assessment. The results confirm that historical firm-level information provides a reliable basis for default prediction, with tree-based models consistently outperforming statistical and neural network approaches. The stacked ensemble achieved the strongest overall performance, whereas C5.0 and CHAID showed that interpretable classifiers can also deliver competitive predictive accuracy. A champion–challenger deployment architecture is proposed, in which the ensemble serves as the performance-oriented champion and interpretable models act as challengers; this arrangement contributes to the operational resilience of the credit-risk assessment process and aligns with macroprudential expectations of model governance, auditability, and explainability. The study offers a replicable methodological framework for integrating data-driven decision support into credit evaluation in comparable banking settings. Full article

Autism Spectrum Disorder is a long-term neurodevelopmental disorder. Early diagnosis is crucial for timely rehabilitation and intervention. Recently, machine learning and deep learning techniques have been widely explored and have produced encouraging results using eye-tracking scanpath images for the early detection of ASD. However, these approaches exhibit inconsistent performance and classification error rates, as well as limited generalization, due to differences in learning approaches and architectural designs across individual models. To address these problems, we employed a weighted-average ensemble of deep learning models using eye-tracking scanpath images. In this work, two different pretrained convolutional neural networks were selected, including Xception and VGG16, based on their proven efficacy and performance. Moreover, we fine-tuned each model individually and evaluated them on a dataset containing eye-tracking scanpath images. We implemented a weighted-average ensemble technique to combine the diverse predictions of the two models. It reduces classification errors and improves the model’s generalization and overall performance. The adopted weighted-average ensemble technique achieved an accuracy of 98.18%, with a perfect recall, and a competitive Area Under the Curve (AUC) of 99.59%. These findings highlight that applying a weighted average to integrate multiple models’ predictions strengthens the generalization and reliability of ASD detection. Full article

Cybersecurity insider threats remain a significant challenge for modern organisations due to their potential to cause substantial financial and reputational damage. This paper presents a systematic review of insider-threat research (2019–2026) using the PRISMA methodology and introduces an empirically validated ensemble framework for insider-threat detection. The proposed approach combines User-Based Sequences (UBS), a self-supervised Transformer trained on next-token prediction and time-gap modelling, and an unsupervised anomaly detection ensemble operating on model-derived behavioural features. An answers directory is incorporated to provide grounded truth for insider entities and episodes within the CERT r6.2 dataset, enabling direct validation of detection outcomes. The framework integrates behavioural theory with machine-learning techniques to improve understanding of insider-threat precursors. Evaluation was performed using a seven-stage Isolation Forest ensemble incorporating multimodal behavioural and technical data streams. The approach successfully identified all insider users, achieving 100% recall and an AUROC of 0.93. Comparative analysis against a previously reported model showed comparable AUROC and perfect recall despite differences in evaluation methodology. While precision remained low (0.004) due to the extreme class imbalance in the full CERT r6.2 population (5 insiders among 4000 users), the results highlight the operational challenges of insider-threat detection in realistic enterprise environments. This research contributes a novel, reproducible framework that combines behavioural theory and advanced machine learning to support the detection and analysis of insider threats. Full article

In casting processes, mechanical properties such as hardness are highly sensitive to both chemical composition and process parameters, making parameter design a complex and uncertain task during the embodiment stage of engineering design. Conventional trial-and-error-based approaches are often costly, time-consuming, and impractical in industrial environments. To address these challenges, this study proposes an optimized fuzzy artificial neural network (FANN)-based decision-support approach for hardness-oriented parameter design in a casting process. The developed model uses chemical composition variables, including carbon, silicon, manganese, phosphorus, sulfur, chromium, copper, and tin, together with process parameters such as casting temperature and casting time as inputs, while Brinell hardness is considered as the output. A dataset consisting of 170 experimental casting samples was employed; 128 samples were used for model development and hyperparameter selection, and 42 samples were reserved as an independent final test set. The proposed model was implemented as a scaled direct FANN weighted ensemble, in which fuzzified input variables were used to predict standardized continuous hardness values. A total of 300 FANN configurations were evaluated using five-fold cross-validation, and the five best-performing configurations were combined through RMSE-based weighted ensemble averaging. The final model was compared with Random Forest, Linear Regression, Ridge Regression, and SVR-RBF models using MSE, RMSE, MAE, R², MAPE, normalized RMSE, and ±5% prediction success rate. The results showed that the optimized FANN ensemble achieved the lowest mean RMSE in the full-data five-fold cross-validation analysis, slightly outperforming the Random Forest benchmark. In the independent final test set, Random Forest produced the lowest prediction error, whereas the proposed FANN ensemble remained competitive and achieved the same ±5% prediction success rate as Random Forest, Linear Regression, and Ridge Regression. Furthermore, a target-hardness case study demonstrated that the proposed approach could identify candidate casting conditions very close to a desired hardness level, with the nearest prediction reaching 202.985 HB for a target value of 203 HB. These findings indicate that the proposed FANN-based framework can serve not only as a hardness prediction model but also as a practical fuzzy decision-support tool for target-hardness-oriented parameter design in casting processes. Full article

Deep learning models for power quality (PQ) disturbance classification remain critically vulnerable to adversarial perturbations, with classification performance degrading severely under white-box attacks. Existing defenses address individual models in isolation and provide no mechanism for operators to assess whether the system is under attack or which classifier remains trustworthy. This paper proposes a two-stage framework that combines adversarial training with large language model (LLM) reasoning to improve both robustness and interpretability. In the first stage, four architecturally diverse classifiers, including a proposed Multi-Scale Temporal Attention Network (MSTAN), are evaluated under four adversarial attacks (FGSM, PGD, C&W, and UAP), and their failure patterns are recorded as structured vulnerability fingerprints. The ensemble is then retrained via adversarial training on mixed clean and perturbed signals. In the second stage, an LLM analyzes the ensemble predictions alongside the fingerprint knowledge base to perform attack detection, fingerprint-guided meta-classification, and operator-facing threat report generation. On a 17-class, 255,000-signal synthetic benchmark, adversarial training recovers FGSM and PGD accuracy from below 25% to the 53–78% range, with MSTAN achieving the highest post-training robustness (78.26% under FGSM, 65.41% under PGD). The LLM reasoning layer provides an additional 3.5–6.2 percentage point improvement over majority voting by selecting the most reliable ensemble member based on the inferred attack condition, and detects adversarial attacks with 87.6% overall accuracy. To our knowledge, this is the first integration of LLM-based ensemble reasoning into the PQ adversarial robustness pipeline and the first application of the C&W optimization attack to power quality signals. Full article

The prediction of student engagement characteristics involves forecasting and analyzing student interaction with educational materials using engagement prediction models. This process encompasses the prediction of cognitive, behavioral, and emotional dimensions of engagement. The existing student engagement prediction models have some limitations, including poor convergence, less generalizability, complexity issues, overfitting, false errors, and limited resources. Hence, the research proposes the Multi-Instance Localization-based Gradient Boosted Long Short-Term Memory (MIL-GBLTM) model to tackle the challenge of predicting student engagement characteristics in online classes. The integration of effective MIL with a Triplet Attention mechanism focuses on the significant features that help with engagement prediction; LSTM layers capture intricate sequential patterns, and fractional gradient boosting is used for fine-tuning for accurate prediction, alongside ensemble-based learning. The LSTM layers with the Triplet Attention module refine temporal attention, and Fractional Gradient Boosting ensures the model’s adaptability and robustness. By combining these components, the proposed model is able to predict accurate student engagement with high convergence. This integrated approach enhances the capabilities of engagement prediction models in educational contexts, facilitating more effective interventions and personalized student support in online learning environments. Experimental results demonstrate that the proposed MIL-GBLTM model outperforms other existing models by achieving the highest accuracy of 96.55% with a k-fold of 10, utilizing the wacv2016 dataset. Full article

Cathaya argyrophylla Chun et Kuang is an endangered relict gymnosperm endemic to China. Its habitat has been severely fragmented due to Quaternary glaciations, a condition further exacerbated by modern, fragmented administrative management. We compiled 98 spatially filtered occurrence records across four provinces and developed a combined analysis framework integrating the Biomod2 ensemble model with the Marxan systematic planning algorithm. Our optimal model (TSS = 0.911, AUC = 0.986) identified mean diurnal range and ultraviolet-B seasonality radiation as the dominant ecophysiological drivers of the species’ distribution. Currently, suitable habitats cover 7.10% of the study area, with highly suitable habitats accounting for only 3.08% (21.76 × 10³ km²). Priority conservation areas account for 2.48% (17.55 × 10³ km²) of the total area. A gap analysis revealed that 76.98% (13.51 × 10³ km²) of the optimized priority conservation areas currently lack formal protection under China’s protected area system and the World Database on Protected Areas. Under four future climate scenarios (2030s–2090s), projections indicated overall habitat contraction, with limited spatial expansion observed only under specific scenarios (SSP1-2.6 in the 2030s and 2090s; SSP5-8.5 in the 2030s), and the population centroid was projected to shift southeastward by an average of 42.67 km in Huaihua City. Twenty-one core habitat patches were identified under current climate conditions. As these core habitat patches are concentrated along interprovincial boundaries, specifically the Dalou Mountains and the Yuecheng Ridge, our findings emphasize the need to bridge local administrative barriers. This spatial framework provides actionable guidelines for establishing transboundary protected areas, optimizing in situ conservation networks, and implementing model-based assisted migration. Full article

Reliable estimates of future precipitation are essential for adaptive water management in large river basins. This study presents a machine-learning approach that combines six CMIP6 models to examine precipitation changes in the Yangtze River Basin. ERA5 monthly precipitation for 1979–2025 served as the reanalysis reference. The random forest model incorporated individual model outputs, ensemble statistics, geographic variables, and monthly cyclic terms. It was trained with data from 1979–2009, evaluated for 2010–2014, and then applied to the period 2015–2099 under SSP1-2.6, SSP2-4.5, and SSP5-8.5. Compared with the simple multi-model mean, the proposed method showed better agreement with ERA5 and generally smaller reconstruction errors during the validation period. Annual precipitation is projected to increase under all three pathways, with the largest increase under SSP5-8.5. Precipitation remains concentrated from May to August, while spring totals and intra-annual variability increase more clearly under high-emission conditions. Mean precipitation remains highest in the humid middle and lower reaches, while the magnitude and significance of future trends vary across the basin. Inter-model spread remains greater than the differences among emission pathways and reaches 85.92 mm under SSP5-8.5 during 2071–2099. These results represent uncertainty-aware climate estimates rather than verified forecasts. They can support flood-risk assessment, reservoir planning, and adaptive water management in the Yangtze River Basin. Full article

Diffusion models for 3D genome structures make inference an ensemble-generation and tiling problem. In the released ChromoGen workflow, millions of independent denoising trajectories are executed through a single-GPU path, while overlapping genomic windows are sampled without enforcing consistency of their shared physical interval. We introduce ParaChromo, a parallel inference framework for conditioned, tiled 3D genome diffusion workloads built around the trained diffusion U-Net and distance-map interface. ParaChromo organizes the workload into three inference-layer modules: a workload-dispatch module schedules region, guidance, and sample chunks across worker groups; an encoder-aware sharded-conditioning module scales and shards the EPCOT front end with FSDP while keeping the inner-loop U-Net replicated; and a seam-coherent tiled-synchronization module projects the shared 12-bead overlap of adjacent reverse chains in distance-map space. On eight A6000 GPUs, the combined reduced-step and task-parallel systems path raises throughput from $2.356\pm 0.003$ to $235.71\pm 1.120$ samples/s, a $100.04\pm 0.486$-fold gain over the released single-GPU baseline. The reduced-step setting is supported by a sweep from 50 to 1000 DDIM steps, where distance-distribution and Hi-C-based metrics remain stable across four chromosomes. For the synchronization module, the chr22 seam discrepancy falls from 150.9 pm to 7.9 pm, while matched internal and Hi-C-based quality metrics are preserved. The synchronized chr22 run also gives a chromosome-scale coordinate rendering over 32 paper-aligned tiles. Together, these results show that conditioned, tiled 3D genome diffusion can be executed as a scalable workload when throughput parallelism, sampler length, encoder placement, and spatial consistency are treated as separate but compatible constraints. Full article

The commercialization of lithium metal batteries, a key technology for high-density energy storage, is hindered by issues with coulombic efficiency, which dictates battery stability and life. In this paper, we propose a machine learning framework to forecast liquid electrolyte efficiency, where two experimental data sources were combined to create a curated dataset of 283 records. In addition, to assess several ensemble learning algorithms, thirteen chemical descriptors were used, as well as interpretability analysis and Bayesian optimization to guarantee physicochemical consistency. We found that the optimized CatBoost model achieved a coefficient of determination (R2) of 0.61 on the test set and a mean squared error (MSE) of 0.0924, representing a significant improvement in predictive accuracy compared to previous standards. Furthermore, these results demonstrate that regulating oxygen levels in solvent environments is a key component of high-density energy storage. These results can serve as a virtual screening tool in order to discover high-performance electrolytes with the minimum experimental costs. Full article

The availability of critical spare parts is essential for maintaining the reliability and operational continuity of combined cycle power plants. However, demand for critical spare parts is typically sparse, intermittent, and highly non-linear, which limits the effectiveness of conventional forecasting approaches based on historical averages or expert judgment. Inaccurate demand estimation may lead to excessive inventory, high holding costs, or stock shortages that increase downtime risks. To address these challenges, this study applies ensemble learning methods to improve demand forecasting accuracy for critical spare parts in a combined cycle power plant. Procurement and usage data from 2020 to 2024 were analyzed using a time-series splitting approach, with model performance assessed using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). To avoid bias caused by zero-demand periods, zero actual values were excluded from MAPE calculations. The results show that the tuned XGBoost model consistently performs better than Random Forest by producing lower forecasting errors and more stable predictions under intermittent demand conditions. These findings indicate that ensemble learning can support more effective procurement planning, inventory control, and maintenance decision-making in combined cycle power plant operations. Full article

Electromagnetic shielding has become a practical concern in buildings and structures exposed to persistent interference. This paper reports experimental measurements of the frequency-dependent shielding properties of a square-resonator-integrated double-concrete structure, using a free-space S-parameter setup built around WR229 waveguide adaptors and horn antennas. Three variables were tested: concrete thickness D, relative permittivity ε_r, and relative magnetic permeability μ_r. Both ε_r and μ_r were characterized experimentally from carbon-fibre- and copper-slag-modified concrete rather than taken from standard tables. The novelty of the study lies in combining experimentally characterized concrete electromagnetic properties, an embedded square-resonator geometry, and explainability-driven machine learning analysis within a single experimental framework for cement-based EMI shielding design. A total of 96 parameter combinations were evaluated using calibrated S₁₁ and reference-corrected S₂₁ responses across 3.3–4.9 GHz. Thickness and electromagnetic material properties interacted—neither governed shielding performance on its own. The strongest transmission attenuation occurred at D = 5, ε_r = 7, and μ_r = 1.2, where minimum S₂₁ reached approximately −62.98 dB at 3.6392 GHz. S₁₁ varied considerably less than S₂₁ across the tested combinations, suggesting transmission suppression is the dominant mechanism rather than reflection enhancement. A machine learning analysis confirmed that nonlinear ensemble models outperformed the linear baseline and identified thickness as the most influential predictor of minimum S₂₁. Full article