We present a comprehensive parameterized complexity framework for the cryptanalysis of classical ciphersystems composed of multiple cryptographic transformation layers. Our main theoretical contributionsare threefold. First, we prove that Cipher-Decode is fixed-parameter tractable (FPT) when parameterizedby the structural complexity vector (k, ℓmax, |Σ|), yielding an algorithm with running time O∗(|Σ|k·ℓmax ).Second, we establish that the problem is para-NP-hard when parameterized solely by the number of layersk, and W[1]-hard via a complete formal reduction from k-Clique, providing a tight complexity dichotomy.Third, we design a branch-and-bound algorithm with statistical pruning based on Kullback-Leibler divergenceand the Index of Coincidence, reducing the effective search space by two to six orders of magnituderelative to exhaustive search. Experimental evaluation on 47 benchmark instances spanning Vigenère,substitution, ADFGVX, and multi-layer composite ciphers confirms the theoretical scaling predictions(R2 = 0.9998) and demonstrates speedups of 115× to over 2000× against naive baselines, with directcomparisons against simulated annealing and beam search. To the best of our knowledge, this is the firstrigorous FPT analysis of multi-layer classical cipher cryptanalysis.
Traditional cryptographic hash functions and existing neural hashing schemes suffer from a fundamental vulnerability: their deterministic nature renders them susceptible to rainbow table attacks and precomputation threats. To address this, this paper presents the Temporal-Stochastic Neural Cipher (TSNC), a novel framework that bridges dynamic security with deterministic verification. Unlike static architectures, TSNC eliminates the computationally expensive training phase by leveraging account creation timestamps and passwords to generate cryptographically secure seeds. The framework establishes a robust encryption pipeline through three logically cascaded modules: (1) timestamp-driven dynamic dictionary construction that ensures temporal uniqueness for character embeddings; (2) sequential feature extraction utilizing recurrent transformations to capture inter-character dependencies; and (3) a stochastic neural network with input-dependent dynamic topology, where the network depth and width adaptively evolve based on intermediate features. Theoretical analysis confirms the method's computational irreversibility and intractability against reverse engineering. Comprehensive simulations validate TSNC's security properties, demonstrating strict collision resistance (zero collisions in $10^7$ trials), a profound avalanche effect with output feature vectors exhibiting near-zero cosine similarity under minimal input perturbations, and temporal orthogonality. Furthermore, the algorithm incorporates a constant-time execution mechanism, effectively mitigating timing side-channel attacks while maintaining high computational efficiency. To promote reproducibility and further research, the source code of the proposed algorithm is publicly available at https://github.com/yongqianxiao/TSNC.
Convolutional Neural Networks (CNNs) have achieved significant performance in various computer visiontasks, but at the cost of enormous computing resources. To alleviate this dilemma, this paper proposes alightweight and efficient multiscale attention fusion module (SCEMA) by combining spatial, and channelreconstruction convolution with an efficient attention mechanism and adds the module to the YOLOV8 networkstructure named YOLO-SCEMA. SCEMA adopts a parallel processing strategy, with the left branchperforming feature refinement operations through spatial and channel reconstruction units to reduce redundantcalculations, and the right branch effectively integrating features of different scales through featuregrouping and cross-spatial learning, enhancing the model’s understanding of multiscale image contentand improving its performance in handling complex image structures. The experimental results on the opensource datasets ExDark,VisDrone2019 and FYP show that YOLO-SCEMA has increased the mAP(50) scoreby 7.37%,3.24% and 1.5% , and compared to the YOLOv8 benchmark while reducing the parameter andcomputational complexity by 36.9% and 8.6%, respectively. Compared with the latest YOLO series, YOLOSCEMAperforms better in detection accuracy and parameter quantity.
Genetic Programming (GP) can be used to evolve human-interpretable Behavior Tree controllers for realtimestrategy games. Current Behavior Trees-based GP approaches compute fitness only at the end ofeach episode, which does not allow learning from finer-grained tactical decisions during gameplay. In thiswork, we integrate tabular Q-learning within the BT controller to gate actions at terminal nodes based onlearned action-values, and collect additional rewards at each tick of the game that are used to augmentthe final fitness signal. Since the Q-table is reset at the start of each episode, our approach allows thelearned information to affect fitness, while keeping the learned values from one generation separate fromsubsequent generations (a form of Baldwinian learning). Applied to plain-terrain MicroRTS micromanagementchallenge (population of 100 BTs, 2000 generations, against a deterministic rush opponent), ourapproach achieves a maximum fitness of 26.44 compared to GP’s max fitness of 22.0, a relative improvementof 20.2%. Mean best-fitness was also increased by 17.8% compared to the GP-only baseline. Theseresults are reported from a single representative run per configuration; multi-seed replication is identifiedas a priority for future work. The additional per-episode RL reward signal is strongly correlated with elitefitness (r = 0.89), confirming that it provides informative guidance for the evaluation. Unit coordinationexhibited by the hybrid agents is also found to be more structured, with defined front-line and support roles.The agent maintains its interpretable Behavior Tree structure.
Lithium-ion batteries have been widely utilized in modern society due to their excellent performance, making the accurate prediction of their remaining useful life (RUL) of paramount importance. Current RUL prediction methods are primarily categorized into battery discharge model-based approaches and data-driven approaches. However, the charging and discharging processes of batteries are typically accompanied by nonlinear variations in internal structural parameters, which pose significant challenges for model-based prediction methods. To address this issue, this paper proposes a predictive model based on Long Short-Term Memory (LSTM) networks optimized by an Improved Dream Optimization Algorithm (IDOA). First, to overcome the limitations of the original Dream Optimization Algorithm (DOA)—namely, the tendencies to become trapped in local optima and slow convergence—an optimal point set is employed for population initialization, and an adaptive population reduction mechanism is introduced to accelerate convergence. Second, the IDOA is utilized to optimize the hyperparameters of the LSTM network, thereby enhancing its predictive capability. Finally, simulation experiments are conducted using the CALCE and NASA battery datasets via the leave-one-out cross-validation method. The performance is evaluated using the average Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) calculated over ten independent runs. Comparative analysis with the original DOA and other baseline models demonstrates that the proposed IDOA-LSTM model can accurately predict the RUL of lithium-ion batteries with higher precision. Specifically, for the four batteries in the CALCE dataset, the RMSE values are 0.0201, 0.0198, 0.0213, and 0.0186, while the MAE values are 0.0181, 0.0174, 0.0163, and 0.0154, respectively. The results indicate that the average prediction accuracy of the proposed model is improved by 53.8% compared to the unoptimized model.
The explosive expansion of Low Earth Orbit (LEO) satellite mega-constellations presents critical challenges in operational risk management, real-time task allocation, and dynamic resource management due to their inherent time-varying topology and systemic uncertainties. To overcome the limitations of traditional optimization methods, this paper proposes a novel, hybrid artificial intelligence framework integrating K-means Clustering, Fuzzy Logic (FL), and Deep Reinforcement Learning for Task Allocation (DRL-TA). First, an adaptive K-means clustering mechanism periodically segments the constellation based on real-time channel quality and traffic load, effectively reducing network state dimensionality and overhead for the DRL agent. Second, a Fuzzy Inference System is employed to model non-deterministic operational elements (e.g., temperature, orbital deviation) and predict the real-time Safety Risk Score 〖Fuzzy〗_risk. This interpretable risk score is integrated as a penalty term within the DRL agent's reward function. Finally, the DRL-TA algorithm learns the optimal policy for computationally offloading and resource allocation by jointly analyzing the clustered network state and the fuzzy-predicted risk. Validated on a simulated 1,000-satellite LEO constellation over 10,000 training episodes, the integrated DRL-TA framework demonstrates significant performance gains: achieving a 25% reduction in average task completion delay and a 15% improvement in overall task success rate compared to conventional load-balancing and pure DRL baseline methods. The DRL policy exhibited stable convergence within 8,500 episodes, with a final average episodic return accuracy exceeding 97% of the theoretical maximum. This demonstrates the framework's efficacy in creating a reliable, high-performance, and risk-aware LEO edge computing environment.
In power system intelligent fault detection, real-time monitoring is critical due to grid complexity. To address the challenges posed by the complexity of continuous action selection and gradient estimation errors in transmission line fault monitoring, we developed a Deep Deterministic Policy Gradient (DDPG) algorithm enhanced by the Stochastic Variance Reduction Gradient (SVRG) method, termed SVRG-DDPG. This algorithm leverages the SVRG technique to mitigate the gradient estimation errors typically encountered in the DDPG algorithm. Utilizing the SVRG-DDPG framework, we further developed a transmission line fault monitoring model that directly employs real-time voltage sensor data from actual transmission line environments, encompassing a range of state information from normal operation to fault conditions. To achieve real-time monitoring and optimize system performance, we also propose a transmission line fault monitoring system based on a lightweight edge computing architecture. Using real-time voltage sensor data, the SVRG-DDPG-based fault monitoring model achieves a residual error within 30 kV accuracy. To enable real-time fault diagnosis in resource-constrained edge environments, we propose a lightweight edge-cloud collaborative architecture that dynamically allocates computational resources based on fault severity and sensor data volume. The framework is validated using high-fidelity simulation data from a power grid (covering 8 fault types under 12 operational conditions, e.g., humidity >90%, load fluctuations ±40%), which aligns with the dynamic resource demands of edge devices in practical systems. Finally, our method achieves 92% accuracy in fault diagnosis with 42% lower latency compared to baselines, leveraging SVRG-enhanced DDPG for adaptive edge-cloud synchronization. Experimental results on real-world 5G-V2X data validate its suitability for low-latency transmission line monitoring.
With the global prevalence of tomato diseases causing 20 to 40% annual crop losses and over USD 220 billion in economic damage, traditional manual scouting and laboratory diagnostics prove labor intensive, subjective, delayed, and impractical for resource constrained rural farmers. To address this challenge, this study proposes a lightweight 17 layer convolutional neural network (CNN) model enhanced by comprehensive data augmentation, effectively classifying nine prevalent tomato leaf diseases Bacterial Spot, Early Blight, Late Blight, Leaf Mold, Septoria Leaf Spot, Spider Mites, Target Spot, Yellow Leaf Curl Virus, and Mosaic Virus using the PlantVillage dataset of 16,012 images. The experiment utilized 80/20 train test splits with Adam optimizer (learning rate 0.001), categorical cross entropy loss, 50 epochs, and batch size 32. The proposed CNN was compared with pretrained InceptionV3 and ResNet152V2 baselines. Experimental results demonstrate the model achieves state of the art performance with 95.28% test accuracy, 97.80% training accuracy, 0.970 macro F1 score, 0.932 micro MCC, and 0.983 micro average AUC, outperforming InceptionV3 (81.54%) and ResNet152V2 (85.89%) by 13.74% and 9.39% respectively, while surpassing tomato specific SOTA VGG 19 (93%). Ablation experiments confirm augmentation yields 16.68% accuracy improvement over non augmented baselines. The model powers a React Native Android app with TensorFlow Lite INT8 quantization (7.1 MB), delivering sub 200 ms inference for online cloud analysis via FastAPI and offline edge computing, providing farmers real time diagnostics with robust generalization across diverse field conditions and significant practical value for precision agriculture and food security.
To address the issues of cross-domain data privacy protection and collaborative analysis in the detection of Advanced Persistent Threat (APT) attack chains, this paper proposes an APT attack collaborative detection and privacy enhancement method based on hybrid federated learning. This method constructs a two-layer federated learning framework of horizontal cross-organization collaboration and vertical multi-feature fusion, realizing the deep fusion of multi-source threat intelligence under the premise of privacy protection. By designing an adaptive differential privacy mechanism and dynamically optimizing the noise addition strategy, it minimizes the loss of model performance while guaranteeing data security. At the same time, this paper introduces a time-series graph neural network detection model to achieve accurate perception and correlation analysis of multi-stage behaviors of APT attacks. The experimental results demonstrate that HybridFL-APT achieves a macro-averaged F1-score of 0.914 and an attack stage identification accuracy of 0.867. Compared to the standard FedAvg algorithm, the proposed framework reduces the cumulative communication overhead by 30.9% while maintaining robust privacy protection even at a strict privacy budget (ε=0.1).
The dramatic changes in climate, the rising water shortage, and the rising demands of agricultural productivity have resulted in irrigation management as a burning issue in contemporary agriculture. In an attempt to curb these drawbacks, this paper introduces a combined Artificial Intelligence (AI) and Internet of Things (IoT)-based smart agriculture system to manage precision in controlling irrigation measures. The suggested design uses IoT sensors as a continuous monitoring of real-time environmental and soil conditions such as the moisture content of the soil, ambient temperature, humidity, and climatic conditions. These data are sent to a central processing unit and machine learning models are used to find the complex and dynamic patterns that enable a good prediction of the crop water needs. In order to improve the accuracy of the prediction and the efficiency of the system, the Improved Seagull Optimization Algorithm-Adaptive Support Vector Machine (ISOA-ASVM) model is introduced. The optimization algorithm is useful to tune the SVM parameters, enhancing the performance of generalization and minimizing the computational cost. The model is trained and tested using a dataset of 3000 records of the IoT-based sensor on agricultural fields that consisted of the main environmental and soil characteristics such as soil moisture, temperature, humidity, rainfall, light intensity, and soil pH. Minimum performance of the maximum normalization and feature selection are used to increase model stability, and generalization. The evaluation criterion is 10-fold cross-validation and performance is compared to that of the Random Forest, Naïve Bayes and KNN classifiers. The proposed ISOA-ASVM has a high predictive power and strong robustness with high accuracy of 99.3, precision of 96.1, recall of 97.6 and F1-score of 98.6. Low variance and consistent cross fold performance is confirmed by statistical analysis. The acquired AI-IoT system will facilitate automated control of irrigation, remote monitoring, and real-time decision-making and reduce the involvement of humans and operational expenses. The results affirm that the combination of smart machine learning models and IoT sensing infrastructure can be of great help in enhancing the water-use efficiency, crop productivity, and sustainability. This paper offers a scalable and efficient solution to smart irrigation systems and helps to build climate-resilient and resource-efficient smart agriculture.
As Big Data systems increasingly shape decision-making across healthcare, finance, hiring, criminal justice, and public governance, concerns regarding algorithmic bias and unfair outcomes have become more pronounced. While data-driven systems can improve efficiency, prediction, and service delivery, they may also reproduce historical inequalities through biased datasets, proxy variables, flawed labels, or opaque optimization processes.This study presents a systematic review and comparative synthesis of bias mitigation strategies in Big Data systems, with the aim of identifying effective approaches for fair algorithm design. Guided by a PRISMA-informed review methodology, literature published between 2018 and 2026 was screened across major academic databases, resulting in 27 studies included for final analysis.The findings classify mitigation approaches into pre-processing, in-processing, post-processing, and hybrid methods. Pre-processing techniques were found effective for addressing representation imbalance and data quality issues, while in-processing methods provided stronger fairness control where model retraining was feasible. Post-processing approaches were most practical for legacy or proprietary systems, whereas hybrid strategies were strongest in high-risk contexts requiring layered safeguards. The review further shows that no single fairness metric or mitigation technique is universally optimal; effectiveness depends on domain risk, bias source, regulatory obligations, and operational constraints.Based on these findings, the paper proposes the Fair Algorithm Design Decision Framework to guide organizations in selecting context-appropriate fairness interventions. The study concludes that bias mitigation should be treated as a continuous lifecycle responsibility integrating data governance, model design, human oversight, and ongoing monitoring.
The paper discloses the analysis of K-means clustering based cooperative spectrum sensing using energy-based detector for Rayleigh fading channel under low SNR (0 dB to -25 dB) using ROC curves with different rules for fusion “OR-OR, OR-AND, AND-OR and AND-AND fusion rules”. As the current communication systems such as IEEE 802.22 WRAN need to operate with better performance parameters in low SNR conditions, therefore our proposed system considers the similar practical environment. In the proposed system, K-means clustering is used to cluster the secondary users with secondary user closest to centroid selected as cluster head. The clustering technique has provided 70-85% of data overhead reduction in comparison to cooperative or collaborative spectrum sensing. The probability-of-detection doubles in comparison to non-cooperative spectrum sensing at 0.1 probability-of-false-alarm. Further, the comparative analysis is done with other clustering techniques (DBSCAN and LEACH) in terms of data overhead, latency and probability-of-detection. The proposed system has provided an advantage of reduced data overhead, deterministic cluster selection and better performance in terms of detection with OR-AND rule-based fusion in low SNR conditions in comparison to other clustering techniques. Further, the performance variation is also analysed with respect to number of clusters and it has been observed that the detection efficacy of OR-OR and AND-AND rule-based fusion remains unaffected by the variation in number of clusters. However, the detection capability of OR-AND/AND-OR rule-based fusion decreases/increases with the increase in clusters. Further, the OR-AND rule-based fusion performed better than AND-OR rule-based fusion till a certain number of clusters in the network and if clusters increase further, AND-OR rule provided superior performance compared to OR-AND rule. All the simulations were conducted in MATLAB 2024b using TCP-IPv6 model assumptions.
Cardiovascular diseases (CVDs), including chronic heart failure (CHF), represent a major global health challenge. Early detection is essential but often limited by data scarcity. This paper explores two key contributions to address this issue: detection of CHF decompensation using phonocardiogram (PCG) recordings and machine learning, and PCGmix, a novel data-augmentation technique tailored to heart sounds. In a study with 37 CHF patients, our models classified decompensated vs. recompensated states with up to 72\% accuracy. PCGmix further improves diagnostic performance in scenarios with limited training data, achieving comparable accuracy to models trained on datasets up to 50\% larger without augmentation.
Breast cancer is a significant cause of cancer-related deaths among women worldwide. Its early identification and screening are essential for improved patient outcomes and reduced mortality rates. Histopathological image analysis is considered as the gold standard for the diagnosis and prognosis of breast cancer. Nevertheless, the complexity of Whole-Slide Images (WSI) and their manual examination make this task time consuming, and prone to pathologist subjectivity. Recently, Deep Learning (DL) technology has achieved remarkable success in computer vision. However, their application still faces critical challenges in pathology analysis, including Region-of-Interest (RoI) scale variations, inter- and intra-class heterogeneity, diverse staining protocols, and the scarcity of annotated datasets. Furthermore, DL model’s findings are opaque and lack decision-level transparency. This study proposes a novel explainable multi-model DL framework for breast cancer classification leveraging histopathological images. The framework integrates Contrast Limited Adaptive Histogram Equalization (CLAHE) for image contrast enhancement, and diverse data augmentation to mitigate class imbalance and overfitting. Proposed architecture ensembles two branches, one employes DenseNet201 benefiting from Transfer Learning (TL) via ImageNet weights, while other utilizes a custom light weight attention based Hierarchal Feature Fusion (HFF) Network. DenseNet201 utilizes multilevel features to effectively tackle gradient vanishing issues and capture intricate feature representations, while HFF-Net, designed specifically for biomedical imaging, leverages HFF stem and multiscale feature extraction with Swish activation to enhance learning stability. Attention mechanism introduced within HFF-Net further refines the output features. Final feature vectors from the two branches are fused at the Global Average Pooling (GAP) layer, consolidating discriminative information. Experimental results on BRACS dataset demonstrate the proposed framework achieves 97.15% accuracy, 92.59% precision, 93.81% recall, and a 93.19% F1-score in screening tasks, while for grading tasks, it attains 84.08% accuracy, 83.39% precision, 83.64% recall, and 83.44% F1-score on 4391 test samples. Additionally, Gradient-Class Activation Mapping (Grad-CAM) saliency heatmap are generated for visual representation of proposed model’s choices, thereby increased transparency. The integration of these advanced techniques significantly enhances diagnostic reliability, addressing the challenges in histopathological image analysis.
Background: Student mental health has become a global concern due to increasing academic stress, anxiety, and depression, which negatively impact learning outcomes, emotional well-being, and social interactions. Traditional assessment methods, such as static questionnaires and periodic counseling, fail to capture the dynamic nature of psychological health. Objective: This research aims to develop an intelligent deep learning (DL) framework using neural networks to predict early mental health risks and generate personalized psychological intervention strategies based on students’ evolving emotional states. Methods: The Student Mental Health and Intervention Dataset, comprising 1,000 student records collected through self-reported surveys, physiological indicators were utilized. Data preprocessing involved normalization, sentiment classification using BERT-based text embeddings, and feature extraction with Convolutional Neural Network (CNN) layers. The integrated features were input into the Octopus Optimization with Attention-Based Memory Network (2OAMN). The Bi-LSTM captures both past and future dependencies in mental health data, enhancing prediction accuracy. The attention mechanism prioritizes the most relevant emotional and behavioral signals, ensuring timely interventions. OOA optimizes model parameters by balancing exploration and exploitation, improving adaptability and predictive accuracy. The framework was implemented in Python using DL libraries to predict mental health risks and generate adaptive intervention recommendations. Results: Experimental evaluation shows that 2OAMN achieves 97.86% accuracy, 98.81% precision, 97.64% recall, and 98.22% F1-score. 2OAMN regression performance includes an RMSE of 5.92, MAE of 4.11, and R² of 0.74. Conclusion: The 2OAMN reliably predicts student mental health risks and generates effective personalized interventions, offering a practical and data-driven solution for improving student psychological well-being.
Smart Money Concepts (SMC) and the Inner Circle Trader (ICT) provide a rich vocabulary of market-structure events, but translating these discretionary ideas into robust, cost-aware execution remains challenging. This paper proposes a structure-aware trading system that (i) detects SMC/ICT events on OHLC data, (ii) learns calibrated probabilities of favorable resolution over a fixed horizon, and (iii) maps the probabilistic edge to a risk-aware position via linear programming (LP) with exposure, turnover, and Conditional Value-at-Risk (CVaR) constraints. On XAUUSD H1 data, the hybrid method achieves a cost-inclusive development return of 46.7% with Sharpe 1.18 and maximum drawdown (MDD) 6.8%, outperforming rules-only (38.6%, Sharpe 0.86) and ML-only (34.1%, Sharpe 0.79) baselines. On a held-out validation window, calibration and tail risk improve further, reaching Brier score [32] 0.186, Expected Calibration Error (ECE) 6.1%, ES0.95 0.036, and MDD 6.3%. Ablations indicate that probability calibration and the CVaR constraint contribute materially to risk-adjusted performance, while session conditioning improves robustness across volatility regimes.
Aiming to address the accuracy and interpretability needs of early warning systems for financial crises in listed companies, this paper proposes a hybrid early warning model that integrates an attention mechanism and LightGBM. When traditional financial early warning models deal with high-dimensional and nonlinear financial data, it is often difficult to capture the differential influence of key indicators, and they are sensitive to quasi-imbalanced data. This study combines the attention mechanism and the LightGBM classifier to perform weighted processing on key financial indicators and trains the model to handle large-scale imbalanced data. The experiment uses data from ST companies and non-ST companies in the Chinese A-share market from 2015 to 2020, constructing a panel dataset of 2,560 listed companies. The model's warning accuracy on the test set is 88.67%, with an AUC value of 0.942, which is 3.2%, 4.8%, and 5.7% higher than the accuracy rates achieved by using LightGBM, XGBoost, and random forest models alone. The attention weights enhance the model's interpretability and identify key warning indicators, such as cash flow ratios and return on assets.
As artificial intelligence (AI) continues to evolve, its role in software development becomes increasingly relevant. In agile environments, particularly within the Scrum framework, AI shows the potential to improve accuracy, productivity, and overall efficiency. Several Scrum related challenges, such as effort estimation or resource allocation, could be solved by applying AI. Despite the potential, many organizations still struggle with successful AI adoption, mainly due to limited knowledge about AI capabilities and challenges in selecting appropriate AI techniques. This study explores the current state of AI integration in Scrum processes through a literature review. A literature search identified 305 records, of which 18 were included in review. The study examines:(a) how much AI has impacted Scrum, (b) which AI techniques have been used to enhance Scrum processes, and (c) how AI affects automation, assistance, and enhancement within Scrum. Data extraction and coding were conducted using a content analysis, while the synthesized findings were interpreted and discussed using a narrative synthesis approach. The analysis identified seven types of AI techniques (regression, classification, clustering, natural language processing, neural networks, search algorithms, and topic modeling) applied across seven out of nineteen Scrum processes. Most AI support was observed in the Plan and Estimate phase, particularly in effort estimation, user story creation, and release planning. In contrast, no applications were identified for processes in the Implement or Review and Retrospect phases. The impact of AI was categorized into enhancement, assistance, and automation, with a balanced distribution across the categories. The findings suggest that current AI applications in Scrum are concentrated on structured, data-rich processes.
This study assesses the effectiveness of transformer-based and neural network models for detecting emotions in tweets. Five models are evaluated: two transformer-based frameworks (DistilBERT and ALBERT), two neural network architectures (CNN and 3CNN-3LSTM), and a hybrid model (3CNN- 3LSTM-GloVe 300x). The models are evaluated based on accuracy, precision, recall, and F1-score. The findings indicate that ALBERT attains the maximum accuracy at 86.38%, succeeded by DistilBERT with an accuracy of 84.35%. The 3CNN-3LSTM model exhibits an accuracy of 83.79%, whilst the CNN model demonstrates the lowest performance at 65.37%. The hybrid 3CNN-3LSTM-GloVe 300x model exhibits a performance of 75.61%. The results demonstrate that transformer-based models surpass neural network models in emotion recognition, especially in recognizing subtle emotional expressions. Nonetheless, transformer-based models demonstrate increased computational expenses, highlighting the necessity for optimization in real-time applications. This study enhances the domain of emotion detection by a comparative comparison of diverse models, emphasizing the benefits of transformers while acknowledging the computational difficulties. The results indicate significant implications for marketing, mental health, and digital communication, highlighting the need for further enhancement of transformer models for effective implementation.
This paper presents a two-stage probabilistic stacking ensemble for admission outcome prediction based on calibrated heterogeneous classifiers and linear decision fusion in probability space. The proposed approach combines HistGradientBoosting, ExtraTrees, and RandomForest models to generate posterior class probabilities, while probability calibration is applied to improve the consistency and comparability of probabilistic estimates. To prevent information leakage, the meta-level training space is formed exclusively from calibrated out-of-fold probability representations. At the second stage, linear meta-models are used to aggregate probabilistic outputs and produce the final decision. The method was evaluated on a real-world dataset collected from the admission campaign of Lviv Polytechnic National University. Experimental studies using holdout validation and stratified cross-validation demonstrate that the proposed ensemble achieves high and stable predictive performance while preserving the quality of probabilistic estimates. In particular, the method reached F1-scores up to 0.990 and MCC values up to 0.979 on the holdout test set, together with low LogLoss values. Comparative analysis with baseline classifiers and standard stacking approaches confirms that calibrated probabilistic fusion improves both classification quality and the reliability of posterior probability estimates in practical decision-support tasks.
