The paper presents a hybrid machine learning model for the spatial segmentation of soils by salinity using multispectral satellite data from Sentinel-2 and climate parameters of the ERA5-Land model. The proposed method aims to solve the problem of accurate soil cover segmentation under climate change and high spatial heterogeneity of data. The approach includes the sequential application of unsupervised learning algorithms (K-Means, hierarchical clustering, DBSCAN), the XGBoost model, and a multitasking neural network that performs simultaneous classification and regression. At the first stage, pseudo-labels are formed using K-Means, then a probabilistic assessment of object membership in classes and ensemble voting of clustering algorithms are carried out. The final model is trained on an extended feature space and demonstrates improved results compared to traditional approaches. Experiments on a sample of 33,624 observations (23,536—training sample, 10,088—test sample) showed an increase in the Silhouette Score value from 0.7840 to 0.8156 and a decrease in the Davies–Bouldin Score from 0.3567 to 0.3022. The classification accuracy was 99.99%, with only one error in more than 10,000 test objects. The results confirmed the proposed method’s high efficiency and applicability for remote monitoring, environmental analysis, and sustainable land management.
Rare event detection in large tabular data remains a computationally challenging problem due to class imbalance, heterogeneous feature distributions, and unstable thresholds. Traditional machine learning approaches based on individual models and fixed thresholds often exhibit limited robustness and reproducibility in such settings. This paper proposes a hybrid ensemble framework for rare event detection that integrates heterogeneous machine learning models through threshold-aware probabilistic aggregation. The framework combines gradient-boosted decision trees, regularized linear models, and neural networks, leveraging their complementary inductive biases. To ensure reproducibility and robust performance evaluation under severe class imbalance, a leaky-controlled evaluation protocol is employed, including rootwise summation, probability calibration, and validation-based threshold optimization. The proposed approach is evaluated on a large tabular dataset containing approximately 50,000 observations. Experimental results demonstrate improved rare event detection and robust generalization performance compared to individual baseline models. Explainability is achieved through Shapley Additive Explanations (SHAP)-based attribution analysis and clustering in the explanation space, enabling transparent analysis of ensemble decision-making behavior. The proposed framework represents a general-purpose computational solution for rare event detection and can be applied to a wide range of data-driven decision-making and anomaly detection problems.
This article presents a comprehensive review of short text clustering using state-of-the-art methods: Bidirectional Encoder Representations from Transformers (BERT), Term Frequency-Inverse Document Frequency (TF-IDF), and the novel hybrid method Latent Dirichlet Allocation+BERT+Autoencoder (LDA+BERT+AE). The article begins by outlining the theoretical foundation of each technique and their merits and limitations. BERT is critiqued for its capability to understand word dependence in text, while TF-IDF is lauded for its applicability in terms of importance assessment. The experimental section compares the efficacy of these methods in clustering short texts, with a specific focus on the hybrid LDA+BERT+AE approach. A detailed examination of the LDA-BERT model’s training and validation loss over 200 epochs shows that the loss values start above 1.2 and quickly decrease to around 0.8 within the first 25 epochs, eventually stabilizing at approximately 0.4. The close alignment of these curves suggests the model’s practical learning and generalization capabilities, with minimal overfitting. The study demonstrates that the hybrid LDA+BERT+AE method significantly enhances text clustering quality compared to individual methods. Based on the findings, the study recommends the optimum choice and use of clustering methods for different short texts and natural language processing operations. The applications of these methods in industrial and educational settings where successful text handling and categorization are critical are also addressed. The study ends by emphasizing the importance of holistic handling of short texts for deeper semantic comprehension and effective information retrieval.
This paper proposes an intelligent and explainable ensemble system for predicting as-partate aminotransferase (AST) levels based on routine biochemical and demographic data from the NHANES dataset. The framework integrates robust preprocessing, adaptive feature encoding, and multi-level ensemble learning within a nested cross-validation (5×3) structure to ensure reproducibility and prevent data leakage. Several regression mod-els—including Random Forest, XGBoost, CatBoost, and stacking ensembles—were sys-tematically compared using R², RMSE, MAE, and MAPE metrics. The results show that the Stacking v2 architecture, combining CatBoost, LightGBM, and Ridge meta-regression, achieves the highest predictive accuracy and stability. Explainable AI analysis using SHAP revealed key biochemical and lifestyle factors influencing AST variability. The pro-posed system provides a modular, interpretable, and reproducible foundation for deci-sion-support applications in intelligent healthcare analytics, aligning with the goals of applied system innovation.
This study proposes an interpretable and high-accuracy ensemble learning framework for predicting aspartate aminotransferase (AST) levels using open-access biomedical datasets. Using a structured pipeline of preprocessing, feature selection, and model ensembling, we evaluated a series of regression algorithms including Random Forest, XGBoost, CatBoost, and three stacking architectures. The best-performing ensemble (Stacking_v2) achieved R² = 0.98 and RMSE = 1.23 on the validation set, surpassing conventional and single-model approaches. Feature importance was assessed using SHAP values, mutual information, and correlation analysis, revealing that gamma-glutamyl transferase, ferritin, and anthropometric markers had the greatest predictive impact. The proposed stacking-based model demonstrates excellent generalization, robust calibration, and high interpretability, and can serve as a benchmark for algorithmic evaluation in medical data modeling. The work highlights the effectiveness of ensemble regression and interpretable AI in real-world clinical prediction tasks using routine biomarkers.