While sentiment analysis has advanced significantly, fine-grained sentiment classification such as aspect-based sentiment analysis (ABSA), continues to present challenges. These difficulties primarily stem from data scarcity and the inherent complexities of identifying sentiments specific to different aspects within a text. The recent emergence of generative artificial intelligence (AI), particularly large language models (LLMs), has fundamentally altered the landscape of data management and integration. To better capture nuanced sentiment differences, this paper proposes a novel approach that leverages LLMs to generate “hard negatives” for ABSA. We introduce the sentiment-aware chain-of-thought (SACoT) framework, which emulates a human-like reasoning process to manipulate the sentiment polarity of aspect terms. This is achieved by integrating chain-of-thought (CoT) reasoning with a few-shot prompting strategy, guiding the LLM to produce sentences with the desired sentiment towards specific aspects. Furthermore, we employ an LLM-as-a-Judge framework to refine these generated hard negatives, thereby enhancing their quality for subsequent use in supervised fine-tuning and contrastive learning. Extensive experiments on benchmark datasets demonstrate that our proposed approach, LLM-SCADA (LLM-guided sentiment-context-aspect data augmentation), significantly improves model performance, especially in transitioning from neutral to positive sentiments (NEU → POS). An in-depth ablation study further underscores the distinct contributions of the generated hard negatives, the contrastive learning approach, and the refinement process. These findings highlight the potential of LLMs to create sentiment-shifted samples based on aspect terms, paving the way for more robust and sophisticated techniques in the field of ABSA.
Energy demand in the transportation sector has become a significant concern due to global economic growth, urbanization, and increased industrial production. This study focuses on the transportation energy demand (TED) problem. The primary dataset includes gross domestic product, population size, and vehicle-kilometer data. Nine different forecasting models were developed by optimizing the weight parameters of linear, exponential, and quadratic regression models using optimization algorithms such as the differential evolution algorithm (DEA), gravitational search algorithm, and crow search algorithm. These algorithms have not been used in TED forecasting before. This study aims to assess the effectiveness of forecasting models and improve their accuracy by addressing gaps in the literature. The models’ performance is evaluated using global error measurement techniques, and the results indicate that DEA outperforms other algorithms and is highly successful, as supported by the literature. The quadratic model of the DEA algorithm was the most successful algorithm with the highest R 2 value (0.987540) and the lowest mean absolute percent error value (5.2415%) compared to other methods. These findings are essential for decision-makers seeking to establish sustainable energy policies and provide a fundamental reference point for energy strategies, transportation planning, and policy-making.
Task scheduling intends to map the user-submitted tasks to the available virtual machines (VMs) based on their capacity. However, the prevailing techniques fail to perform the run-time mapping between the forthcoming workload and cloud resources. This research develops the Lapin Hunt algorithm-optimized multi-parameter-enabled task scheduler (LH-MTS) to perform energy-efficient task scheduling. The task scheduling is managed using the multi-parameter task allocation controller, where the LH-MTS algorithm schedules the tasks based on the VM parameters. The scheduler uses the multi-objective parameters in the cloud to establish the maximum throughput, resource utilization, and reliability, while reducing the makespan and delay. The results reveal that the proposed LH-MTS approach attained a high throughput of 99.98 bps, showing an improvement of 7.55% and 8.25% over the existing task scheduling with multi-objective grey wolf optimizer (TSMGWO) and fruitfly hybridized cuckoo search (FHCS) algorithms. In addition, the proposed LH-MTS approach attained the makespan of 586 s, displaying a significant reduction of 46.44 s compared to TSMGWO, and 155.91 s compared to the FHCS algorithm.
Background Due to the complex, real-time, and resource-constrained nature of the industrial Internet of Things (IIoT) network, cyberattack detection is becoming more challenging. It's hard for traditional detection techniques to adapt to the dynamic environment of IIoT. Although signature-based and anomaly-based approaches are utilized for intrusion detection system (IDS) in IIoT, they have limitations in detecting evolving threats and inefficiency in threat management, which proves the novel IDS model's requirement in the IIoT network environment. Consequently, a novel hybrid ILinkNet–SqueezeNet-based attack detection and mitigation model is introduced, which comprises four stages. Methodology In the initial preprocessing stage, an improved synthetic minority oversampling technique is proposed for the class imbalance solution, while min–max normalization is applied for data scaling. Afterward, feature extraction is employed, where features like Mutual Information-based features, ReliefF features, holoentropy features, and Higher-Order Statistics features are extracted to provide accurate attack detection. To choose the most useful features from the extracted features, an improved filter-based feature selection is conducted. With these chosen features, the attack detection process is conducted using the proposed hybrid ILinkNet–SqueezeNet-based model, while an improved entropy-based model is proposed for attack mitigation in IIoT. Results The experimental evaluation demonstrates that the proposed hybrid model significantly outperforms traditional methods in terms of accuracy, sensitivity, specificity, and precision, as evidenced by the experimental results using the WUSTL-IIoT-2018 and Internet of Things Intrusion Datasets. The model achieves detection accuracies of up to 96.6% with a sensitivity of 0.97 and specificity of 0.976, showcasing its effectiveness in providing robust security in IIoT environments. Conclusion These findings highlight the model's potential to enhance IIoT security by accurately detecting and mitigating evolving cyber threats in dynamic and resource-constrained environments.
Blockchain is a decentralized, public, and distributed ledger designed to securely record and track transactions. It possesses the potential to transform various industries, including healthcare, supply chain management, and financial services, by enhancing transparency, efficiency, and trust. Despite its promise, blockchain development continues to face several challenges, particularly concerning security, scalability, and standardization. This paper provides a comprehensive analysis of blockchain technology, focusing on its quality of service (QoS), security mechanisms, and the latest frameworks and models shaping its evolution. Furthermore, it examines existing limitations and identifies key open research challenges that must be addressed for broader adoption. The findings suggest that blockchain can substantially improve operational efficiency and data integrity across multiple domains; however, realizing its full potential requires continued research and technological advancement to overcome current barriers.
Big data classification has become popular for classifying important data such as healthcare, stock market, and other fields’ data for effective handling of many classical databases. In recent years, deep learning and machine learning techniques have become more reliable for such classification, which has resulted in effective results with better accuracy. In this survey, articles from around 10 years are taken for analyzing various techniques recently for big data classification, and this study helps to improve the future evolution of fine, adequate, and highly acceptable models for classification-based approaches. This survey contains the advantages and disadvantages of several methods, which aid in overcoming the challenges of building new ensemble models for classification. Moreover, this research provides complete knowledge about the existing methods and emphasizes the ideas for creating high-potential models as well as paving the way for inventing a real-world application to classify big data. The performance metrics to measure the classification techniques are accuracy, sensitivity, and specificity, and also provide improvement for future classification processes.
This study introduces an optimization-based approach to improve the longevity and energy efficiency of wireless sensor networks (WSNs) while ensuring comprehensive target coverage. The suggested method categorizes sensor nodes into separate coverage sets to decrease duplicate data transmission and lower total energy consumption. During each operational period, a single coverage set is active, and a mixed integer linear programming (MILP) model is employed to optimize the selection of cluster heads and multi-hop routing. The model concurrently addresses clustering, routing, and variable sensing radii to enhance energy distribution among nodes. Simulation results indicate that the suggested strategy markedly prolongs network longevity compared to traditional strategies, such as T-LEACH. The analysis of the influence of differing coverage set quantities on network performance indicates that an optimal number of sets achieves a balance between energy conservation and effective coverage. The research presents an enhanced energy consumption model that incorporates both sensing and communication expenses. In comparison to baseline techniques, the proposed framework attains more consistent energy utilization and facilitates extended network operation without compromising monitoring precision. The amalgamation of coverage-aware scheduling, energy-aware clustering, and efficient routing offers a comprehensive solution for extending the lifespan of WSNs, with applicability in energy-limited monitoring applications.
In recent years, machine learning (ML) models have become the top prediction options in agriculture, cybersecurity, healthcare, and finance, among other fields. ML models have assisted in the study of numerous disorders in medical research. Early diagnosis and health schedule management may prevent severe heart disease in many people. This study uses national government repository data to apply ML models to early heart disease detection. This research also introduces two novel models based on swarm intelligence, called ACOKNN (ant colony optimization with K-nearest neighbors) and ACOWNNBoost (ant colony optimization with weighted K-nearest neighbors), to improve heart disease early detection models. The suggested models pick features using ant colony optimization with traditional k-nearest neighbor and weighted k-nearest neighbor with XGBoost. The suggested models are evaluated with two datasets of varying capacity, and results are compared to various ML techniques based on precision, accuracy, recall, F1 score, and receiver operating characteristic area under the curve score. The investigation shows that the suggested model achieved 92% and 98% accuracy with Dataset-1 and 92.6% and 98.6% with Dataset-2.
This study proposes an improved random forest algorithm for an educational resource course recommendation network based on boundary value for the problems of low accuracy of educational resource recommendation network in universities, including weak ability to deal with boundary value, and poor adaptability in the face of complex educational environments and diversified user needs, the use of this network can be recommended for better educational resources. The research findings demonstrate that the algorithm error value of the improved random forest algorithm is better in the analysis of the matrix, the algorithm has the best algorithmic performance when the number of forests is 300, at the same time in the system test the algorithm can be smooth and safe through the test, the system in the home page resource recommendation and network performance, the improved algorithm test is good, the uploading and running time is less than 1 s, and the memory accounted for less than 40%. Algorithm model threshold at 0.10 and 0.15, the accuracy trend is the same as the threshold 0.05, while the larger the threshold the higher the accuracy. It can be seen that the improvement of the random forest algorithm can improve the accuracy rate of the current course recommendation, and at the same time, it can complete the course recommendation of the current educational resources, which has certain research significance for the research in this direction.
Retrieving and analyzing vulnerability reports remains a critical challenge in cybersecurity, exacerbated by the exponential growth of disclosed vulnerabilities and the increasing complexity of Proof-of-Concept (PoC) reports. Confronted with massive numbers of vulnerability reports, automated tools and models are urgently required to facilitate the understanding and analysis of vulnerability reports and PoC reports, thereby supporting security professionals in filtering similar vulnerabilities and extracting critical vulnerability attributes. Existing methods for vulnerability semantic similarity either rely on rule-based keyword matching (failing to capture contextual nuances) or generic pre-trained language models, leading to suboptimal retrieval performance. To address these gaps, this work focuses on the semantic similarity analysis of vulnerability descriptions using semantic representation learning. To fully exploit the semantic information embedded in vulnerability reports, we propose a task-specific fine-tuned Sentence Transformer based model for calculating vulnerability semantic similarity. Our approach integrates domain-specific knowledge of vulnerability reports into the model training process, enabling it to capture nuanced semantic relationships unique to cybersecurity. On this basis, we further construct an end-to-end vulnerability retrieval system that uses our fine-tuned similarity model with Elasticsearch vector indexing, realizing intelligent retrieval of both vulnerability and PoC reports. Experimental results demonstrate that the proposed model captures domain knowledge more effectively, and the enriched semantic information significantly enhances the effectiveness of vulnerability report retrieval.
The digital transformation in education, driven by platforms such as massive open online courses and virtual learning environments, has significantly broadened access to education globally. This study explores how machine learning (ML) can enhance and predict learning outcomes by identifying and supporting at-risk students through predictive models. The goal is to improve academic achievement by analyzing student profiles and addressing multiple contributing factors. In this research, ML classification models, specifically support vector classification (SVC), were employed along with two optimizers: the Ebola optimization algorithm and the attack-leave optimizer. The aim was to enhance the model's performance and prediction accuracy. The results indicate that in the training phase, the support vector-based ebola optimization (SVEO) model achieved an accuracy of 0.900, demonstrating moderate performance compared to the SVC model, which achieved an accuracy of 0.887. The support vector-based attack-leave optimization (SVAL) model, however, outperformed the others with the highest accuracy of 0.936.
Knowledge tracing (KT) is a core component of intelligent education systems, which aims to predict students' future performance from their historical learning behaviors. Despite the remarkable progress brought by deep learning, existing models often struggle to effectively capture the dynamic evolution of students' knowledge states, particularly the intertwined processes of knowledge-acquisition and -forgetting. To address this challenge, we propose the Dynamic Knowledge Perception for Temporal Knowledge Tracing (DKPKT) model. DKPKT integrates two complementary modules: the interaction perception module employs a dynamic attention mechanism to extract key interaction features that signal shifts in knowledge states, and the knowledge perception module refines these representations by incorporating knowledge-forgetting and knowledge-acquisition factors, which enable more accurate modeling of knowledge dynamics over time. We evaluate DKPKT against several mainstream baseline models on three real-world educational datasets. Experimental results demonstrate that DKPKT achieves superior predictive performance, validating its effectiveness in modeling knowledge states and enhancing the accuracy of student performance prediction.
The effectiveness of group and online learning environments has been widely recognized. Hence, supporting teachers monitoring of group activities in online learning environments is increasingly important. In this study, we propose algorithms for analyzing discussion transitions and their main themes. The first algorithm aims to help teachers quickly identify groups that require intervention. It uses entities and Wikipedia data in the discussion texts to create models that capture continuously changing discussion content over time. The second algorithm aims to instantaneously identify the discussion topic of the group prior to teacher intervention. From the Wikipedia categories obtained regarding discussion transition analysis, the algorithm selects those with the highest probability of being appropriate for the discussion text. To evaluate the effectiveness of the proposed algorithms, we tested them on the W2E dataset, which includes topics comprising multiple events, and an actual discussion dataset. The results confirmed that the algorithms detected the discussion transitions and main themes with high accuracies.
Pretrained transformer models have demonstrated excellent performance on complex tasks. To improve their inference efficiency, recent studies have introduced the multi-exit mechanism, which enables early exiting through multiple intermediate classifiers. However, the deep architectures of pretrained transformers cause severe gradient conflicts during multi-exit fine-tuning, leading to degraded shallow-exit accuracy and reduced early-exit efficiency. To address this issue, we propose Separate Reverse, a multi-exit training strategy specifically designed for pretrained transformer models. The method iteratively integrates reverse iterative optimization and hierarchical knowledge distillation from deeper to shallower exits, maintaining pretrained parameter integrity, enhances the representation capacity of shallow exits, and coordinates gradient updates across exits to achieve a balanced optimization between shallow and deep classifiers. Experiments on multiple GLUE benchmark datasets using BERT demonstrate that our method significantly improves shallow-exit accuracy, maintains main-exit performance, and accelerates inference for simple samples by a large margin.
Detecting mental illness from short social media posts is challenging because these texts are often brief, fragmented, and lack explicit descriptions of the user's mental state. Prior studies using encoder-based models such as BERT show promise but struggle when key contextual information is missing. To address this, we propose a method that augments posts with interpretive sentences generated by MentaLLaMA-chat, a generative model specialized in mental health, and fine-tunes BERT on the augmented dataset. We curated 1,525 Japanese posts containing the word "mental" (in katakana) from X (formerly Twitter) and manually annotated them according to Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition criteria, labeling 557 posts as positive and 968 as negative. Our method improved recall by 2.4 percentage points compared to models trained on the original posts alone, while maintaining comparable accuracy and precision. Shapley Additive Explanations analysis revealed that tokens introduced by the interpretive sentences-including both negative and positive expressions-enhanced the model's ability to identify mental-distress posts. These results demonstrate that generative-model-based text augmentation effectively provides additional context, enabling more accurate detection of mental illness indicators in short, ambiguous social media posts.
Understanding consumer behavior is vital for businesses seeking to personalize services, optimize marketing strategies, and improve customer retention. However, analyzing such behavior at scale presents significant challenges due to the volume, velocity, and variety of data, as well as the need for accurate and interpretable prediction models. Traditional classification methods often fall short when applied to large-scale, high-dimensional behavioral datasets, leading to issues in scalability, accuracy, and real-time processing. To address these limitations, this paper introduces a novel framework for consumer behavior analysis using an Improved Fuzzy Classification with Bagging and MapReduce Coordination (IFCBMC) approach, specifically designed for big data environments. The primary objectives of this research are: (1) to develop a scalable classification model suitable for distributed data processing, (2) to enhance prediction accuracy through fuzzy rule-based learning, and (3) to evaluate the robustness of the proposed model against existing state-of-the-art classifiers. The process begins with data preprocessing, including cleaning and modified normalization, followed by distribution of the data across a MapReduce architecture to manage scale and speed. Extracted features from multiple data partitions (mappers) are aggregated and processed by an enhanced fuzzy rule-based classification model. To improve prediction robustness, a bagging ensemble strategy is applied, where multiple classifiers are trained on different data subsets, and the best-performing models are randomly selected and merged during the reduce phase. The proposed IFCBMC method outperforms all compared models, achieving the highest accuracy of 0.960, significantly surpassing traditional approaches such as LSTM (0.862), LINKNET (0.866), SQUEEZENET (0.858), SVM (0.863), DCNN (0.860), Bi-GRU (0.854), RNN (0.861), and DNN (0.860).
Network virtualization is a crucial facilitator for the swift advancement of 5G networks by providing flexibility, scalability, and optimal resource allocation. Although network function virtualization and network slicing offer on-demand services for various tenants, the difficulty of developing a resilient security architecture in 5G persists. This paper presents a distributed, multi-layered security infrastructure augmented by an AI-driven anomaly detection module to overcome this gap. The framework utilizes various virtual security functions (VSFs) throughout the layers of the 5G architecture, with the AI module facilitating real-time identification of anomalous traffic patterns and zero-day vulnerabilities. This modular design allocates security tasks among VSFs, activating them only as necessary, hence assuring efficiency and resilience. Simulation outcomes indicate that the suggested method markedly enhances detection precision and security resilience while minimizing processing overhead. Moreover, in comparison to traditional centralized approaches, the improved framework demonstrates higher performance for load balancing, coverage, distance, throughput, and attack detection efficacy.
A novel to address these issues, human activity recognition is developing a three-fold deep learning model. Prior to anything else, the collected raw video frames undergo pre-processing. Wiener filtering, video-to-frame conversion, and contrast enhancement based on contrast limited adaptive histogram equalization are among the activities accomplished during this phase. Characteristics including monogenic binary coding, binary pattern of phase congruency, local Gabor transitional pattern, and chessboard median binary pattern are then recovered from the ROI region that was obtained. Using artificial ecosystem customized bald eagle optimization, the best features from the retrieved features would be selected. In the activity categorization phase, the three-fold deep learning model is trained using the chosen optimal features. Three deep learning models are used to model the activity categorization phase: convolutional neural network (CNN), optimized recurrent neural network (optimized RNN), and bidirectional long short-term memory (Bi-LSTM). With ratings of 94.3%, 94.39%, 92.2%, and 94.03% for 60, 70, 80, and 90 learning percentages, accordingly, the suggested model has shown the maximum detection accuracy. The two types of datasets used are Action Recognition Data Set, Human Action Clips, and Segments Dataset for Recognition and Temporal Localization. Real-time performance may suffer from the drawn-out frame-by-frame video processing and feature extraction procedure. In order to demonstrate the effectiveness of the recommended approach, the efficiency of the suggested approach is finally compared to other conventional models.
This study aims to prevent possible credit risks that might endanger a bank's financial health and credit performance by utilizing machine learning approaches to anticipate consumer eligibility for credit cards. Credit cards are a common type of credit instrument offered by banks and other financial institutions around the globe. However, these institutions are always exposed to credit risks, which frequently lead to non-performing credit facilities because of unreliable repayments. Banks have historically relied on traditional scoring algorithms to evaluate applicants’ creditworthiness to reduce these risks, but these models may not always produce reliable findings. By using predictive algorithms, this initiative aims to help banks and financial organizations identify and interact with creditworthy consumers. In this research, extreme gradient boosting classification and Naive Bayes classification were utilized to forecast credit card approval. Furthermore, three innovative metaheuristic algorithms, namely the grasshopper optimization algorithm, stochastic paint optimization, and sooty tern optimization algorithm, were integrated to enhance the efficacy of the XGBC and NBC models. This hybridization resulted in the creation of novel models: XGBC + GOA (XGGO), XGBC + SPO (XGSP), XGBC + STOA (XGST), NBC + GOA (NBGO), NBC + SPO (NBSP), and NBC + STOA (NBST). Among these, the XGST model demonstrated superior performance, achieving an accuracy of 0.923 in the test phase, while the NBC model exhibited the weakest performance with an accuracy metric value of 0.831. In the Test section, the XGGSS model demonstrates the highest performance among all the other models, achieving an accuracy metric value of 0.932. Additionally, this model exhibits the best performance in terms of precision, with a value of 0.935. The proposed models enhance credit risk assessment, optimizing approval processes, automating decisions, and reducing default rates through metaheuristic-optimized machine learning.
