Intelligent Data Analysis: An International Journal

In recent years, Transformer-based models have dominated the field of long-term time series forecasting. However, the quadratic complexity of attention mechanisms makes both training and inference computationally expensive. The SOFTS model has emerged as an efficient alternative, replacing attention mechanisms with the STAR module to preserve linear complexity while achieving performance comparable to or better than competing approaches. The SOFTS model builds on the iTransformer architecture, which marked a significant advancement in long-term time series forecasting. Although neither iTransformer nor SOFTS incorporates positional embeddings, our analysis revealed a clear opportunity to improve forecasting accuracy by introducing them. However, the straightforward inclusion of positional embeddings leads to convergence and generalization issues. To address this, we propose a simple yet effective technique: during training, positional embeddings are randomly omitted in certain forward passes, which reduces instability and helps the model generalize better. We refer to this novel form of using positional embeddings as Learnable Stochastic Positional Embedding. Additionally, we incorporate multiple dropout layers to mitigate overfitting and improve accuracy. These modifications result in SOFTS++, a fast and accurate model that achieves the best performance on at least 10 out of 12 standard benchmark datasets. By maintaining linear complexity and requiring minimal computational resources, SOFTS++ stands out as a capable and resource-efficient method for multivariate long-term forecasting tasks.

Time series classification(TSC) is an important task in time series analysis and has been around for decades in the data-mining and machine learning communities. Traditional supervised learning model requires a large amount of labeled training data, while unsupervised model has insufficient performance for complex time series. In this paper, we propose a semi-supervised training framework based on the Mean Teacher model. By using no more than 10% labeled data and dynamically adjusting the smoothing coefficient, this framework reduces the model's training time and improves the accuracy of multivariate TSC. Aiming at the difficulty of choosing the receptive field (RF) size of Convolutional Neural Networks (CNN), we use multiple prime numbers as kernel sizes for the Omni-Scale CNN (OS_CNN) to ensure that the overlap of RF between different convolution kernels is minimized, which can cover various regions of the time series comprehensively and improve the diversity and effectiveness of feature extraction. Finally, an enhanced self-attention mechanism based on convolution is used to improve the convergence of model training. Experiments demonstrate that the method in this paper improves more than 10% over 7 various state-of-the-art baselines on 42 different datasets.

The Harris Hawks Optimization (HHO) algorithm has recently garnered extensive attention in research and applications. However, it still faces critical challenges, including premature convergence, degradation of swarm diversity, entrapment in local optima for complex problems, difficulty balancing exploration and exploitation, and a performance heavily dependent on parameter tuning, which lacks universal guidelines. This paper proposes an improved Hawks swarm optimizer embedded with a new search strategy called BAHSO to address the challenges. The improved algorithm enhances its exploration capability while maintaining the strong exploitation characteristics by leveraging the directional perturbation mechanism inspired by the beetle antennae search (BAS) algorithm. Specifically, BAS introduces adaptive step-size adjustments and stochastic directional search to help individuals escape local optima. At the same time, a dynamic energy learning coefficient ensures balanced exploration-exploitation trade-offs throughout the hard-soft siege hunting process. Experimental results and comparisons demonstrate that BAHSO achieves superior convergence accuracy and speed, particularly in high-dimensional and multimodal landscapes, with significantly improved solution quality over the best-performing variants. Furthermore, diversity metrics confirm that BAHSO maintains higher swarm diversity in late-stage iterations than other methods, effectively mitigating stagnation issues.

Feature selection represents a complex multi-objective optimization challenge aimed at identifying the optimal subset of features while maintaining high accuracy within the domain of machine learning, a task known for its difficulty. In this study, we devise a cost function that simultaneously optimizes classification accuracy and the selected features through linear weighting. Subsequently, we introduce an enhanced meta-heuristic approach named improved bald eagle search (IBES) algorithm to effectively optimize the designed cost function. IBES incorporates the opposition-based learning, levy flight, and a nonlinear control parameter strategy into the various stages of the conventional BES algorithm, to develop a wrapper-based feature selection method. Additionally, an efficient binary optimization called V -shape transfer function is employed to convert the solution space from continuous to binary. To assess effectiveness, we utilize 24 well-known data sets from the UCI repository and compare the performance of the proposed methods with a selection of established algorithms from the literature in terms of cost value, classification accuracy, computational time, and selected features. The results affirm the superiority of the proposed method across the majority of the tested data sets.

As the typical task in graph analysis paradigm, node classification is to predict the class label for each node in the given graph. To address the challenge of label scarcity, graph contrastive learning (GCL) has emerged as a mainstream approach for unsupervised node representation learning. By training graph neural networks (GNNs) to distinguish between positive and negative sample pairs across different augmented views, GCL enables effective feature learning. However, most view augmentation strategies in existing GCL methods inevitably introduce additional noise, degrading the quality of node representations. Moreover, these methods often overlook the hierarchical community structures inherent in graphs, which may lead to mislabeling closely connected node pairs as negative samples—a contradiction to the graph homogeneity assumption. To tackle these issues, we propose a GCL-based node classification method rooted in structural entropy. Specifically, we leverage an encoding tree constructed by minimizing structural entropy and an edge-reweighted view generated via an attention mechanism as the augmented views for GCL. This design preserves the integrity of the input view's fundamental structural information. Additionally, by integrating the hierarchical community characteristics of the encoding tree, we develop a graph-tree contrastive loss function to enhance the ability of node representations to capture hierarchical community structures. Extensive experiments show that our method is superior to the state-of-the-art node classification methods in terms of effectiveness and robustness.

There are clustering algorithms (such as DBSCAN) that do not group all data into clusters, but identify some data as noise and exclude it from clusters. In the literature there are no dedicated validity measures for this kind of noise-aware clusterings. Applying the standard measures blindly (which seems to happen in the literature) yields misleading results. We revise top performing, established validity measures to cope with the results of this kind of clustering algorithms and demonstrate that such clusterings may require an additional type of validity check, assessing not only the cluster validity (separation and compactness), but also the validity of the distinction between noise and cluster instances. Additionally, we propose a balanced score, that captures both types of validity to get a holistic validity score. All proposed measures are evaluated on artificial data, mimicking the experiments of the extensive review [Arbelaitz O, Gurrutxaga I, Muguerza J et al. 2013]. The encouraging results demonstrate that the noise aware extension of the Silhouette coefficient and the Score function are least influenced by the noise level.

The idea of Pythagorean fuzzy distance metrics (PFDMs) has been used to discuss sundry selection problems. Existing PFDMs often fail to incorporate all three essential parameters of a Pythagorean fuzzy set (PFS), namely; membership degree (MD), non-membership degree (NMD), and hesitation degree (HD), thereby limiting their precision and effectiveness in real-world decision-making situations. To explore this gap, this research presents a novel three-dimensional (3D) weighted distance metric under the Pythagorean fuzzy framework, which integrates MD, NMD, and HD for a more comprehensive representation of imprecision. The proposed PFDM is theoretically validated and it is shown to fulfill the axioms of a distance function. It is then embedded into the technique for order of preference by similarity to ideal solution (TOPSIS) to enhance multi-criteria decision-making (MCDM), particularly in the context of smartphone selection. A comparative analysis against existing PFDMs demonstrates the superior precision and stability of the new approach. Furthermore, a sensitivity analysis of the novel 3D distance model confirms its robustness with respect to changes in criteria weights. This enhanced 3D distance metric provides a more reliable and interpretable tool for decision-makers in decision making fields.

Brain-Computer Interface (BCI) technology offers potential for improving meditation practices via real-time neural feedback. Traditional EEG signal processing often fails to account for temporal and inter-channel relationships in the data. This study addresses the gap by modeling EEG signals using a multivariate auto-regressive (MVAR) approach, capturing both temporal dynamics and inter-channel interactions. Sparsity is introduced using the group least absolute shrinkage and selection operator (GLASSO), reducing volume conduction issues. From the sparse coefficient matrix, brain connectivity features such as average energy value (EV), phase lag value (PLV), mean absolute correlation (MAC) and magnitude squared coherence (MSC) are extracted. Statistical analyses and scatter plots highlight the influence of these features on cognitive states during meditation. EEG data is classified into EM, NM and CO states using decision trees (DT), Gaussian naive Bayes (GNB), k-nearest neighbor (KNN), and a multi-layer feed-forward neural network (MLFFNN). Metrics include precision, recall, F1-score and accuracy. DT achieved the best performance with 97.12% accuracy, 96.12% precision, 97.39% recall and an F1-score of 97.01%. This study enhances the EEG classification of meditation states by adding sparsity to the MVAR model. Future work could focus on real-time applications for feedback-driven meditation enhancement.

Brain Computer Interface (BCI) technology is presented for improving the quality of life for individuals with physical impairments. It is based on different physiological sensors, among which Electroencephalography (EEG) is exploited for capturing and interpreting brain activity. In spite of its benefits, traditional EEG based classification models suffer from high computational complexity and limited accuracy. Accurate classification of Motor Imagery (MI) EEG signals is major for developing robust and automated BCI systems. This work presents a Deep Learning (DL) model that integrates a Convolutional Neural Network (CNN) with a Multi-Scale Attention (MSA) network which provides better EEG signal classification. Initially, the Multiscale Principal Component Analysis (MSPCA) is exploited for pre-processing the noise signals. Then, the Beluga Whale Optimization (BWO) is presented for selecting optimal features. The proposed model considers a MSA-CNN, which combines parallel convolutional layers with varying kernel sizes and a Squeeze-and-Excitation (SE) based attention mechanism for extracting discriminative features. The suggested model is evaluated by the PhysioNet EEG MI dataset, with outcomes highlighting superior classification performance compared to existing methods and achieved better accuracies of 99.1% on PhysioNet and 99.02% on BCI Competition IV-2a. This hybrid model offered a scalable and efficient solution for real-time MI-EEG classification in BCI applications.

The installation of surveillance systems in public or private settings is increasingly integrated with intelligent and real-time video anomaly detection to proactive threat consideration. A framework is proposed for real-time abnormal behavior detection for streaming video data based on Dynamic Graph Neural Networks (DGNNs). Unlike traditional frame-based models, the DGNN is not restricted to a static set of nodes; instead, it treats each frame as a graph, where nodes denote detected physical objects such as people and vehicles. DGNN introduces real-time frame-wise graph updates unlike prior methods with static or slow graph refresh using motion, pose, and interaction cues, supported by a lightweight spatio-temporal GCN optimized for live anomaly detection. The results convey high performance for the model, which recorded an accuracy of 99.65%, along with the corresponding precision, recall, and F1-score. An FNR and FPR of only 0.35% and 0.03%, respectively, highlighted its strong performance in discriminating normal from anomalous behaviors. The ROC and precision-recall AUC scores zoomed close to 1.0000 in almost all categories, thereby attesting to the system’s robustness where even slight stakes are involved. Following this was the fusion of TensorRT optimization in ensuring the real-time processing capability of the model, with an inference latency of fewer than 200 milliseconds per segment. The DGNN system, therefore, represents a scalable, low-latency option that is capable of real-time detection and classification of diverse and complex real-world abnormal behaviors, making it a novel approach toward smart public safety systems with very little dependency on huge annotated datasets.

Foreign Object Debris (FOD) detection is a critical task for ensuring aviation safety, especially on airport runways where small, diverse, and often occluded objects can pose serious threats to the operating aircraft. Existing deep learning methods often struggle with balancing detection accuracy and computational efficiency in such challenging environments. To address this, we propose LiteFODNet, a lightweight and data-efficient deep learning framework tailored for intelligent FOD detection in surveillance imagery. LiteFODNet consists of the following four novel architectural modules; (i) Compact Multi-Scale Pooling (CMSP) integrates atrous convolutions with global context aggregation for fine-grained multi-scale features, (ii) Spatial-Channel Reducer (SCR) uses depthwise separable filtering for efficient spatial downsampling, (iii) Feature Focus Module (FFM) combines global pooling and dual-stage calibration for dynamic channel emphasis, and (iv) Split Path Attention (SPA) independently learns axis-aligned attention for better spatial localization. Together, these components enhance the model’s ability to generalize across small, complex objects while reducing computational burden. Evaluated on three benchmark FOD datasets (FOD-Tiny, FOD-A multiclass, and FOD-A single-class), LiteFODNet achieves 0.8888% higher mAP@50–95 than YOLOv8n while reducing parameters by 16.39%, inference time by 27.77%, and GFLOPs by 3.66%. These results demonstrate that LiteFODNet offers an intelligent, high-performance solution for real-time FOD detection under constrained resources, with strong potential for deployment in aviation safety monitoring systems.

Steel surface defect detection is regarded as a critical component of quality control in intelligent manufacturing, as its effectiveness directly influences product qualification rates and production costs. To address this issue, a precise defect detection model, LDSE-YOLO, is proposed in this study. Conventional spatial attention mechanisms focus solely on spatial features and fail to resolve the limitations posed by the parameter-sharing nature of convolutional kernels. Additionally, traditional feature pyramid networks lack effective multi-scale contextual modeling, while existing attention mechanisms are often restricted to a single domain, making it difficult to achieve robust object representation and background suppression under complex conditions.To this end, a Local Dynamic Convolution module (LDConv) is first introduced. Unlike static convolutions with fixed patterns, LDConv employs a dynamic weight allocation mechanism to enhance the representation of fine-grained defects. Next, a Spatial-Context Attention Module (SCAM) is proposed, which integrates dilated convolution and adaptive spatial attention to construct a feature pyramid with improved multi-scale perception. This design combines large receptive field feature extraction with dual spatial-channel attention to effectively decouple defect features from background noise in texture-rich environments.Furthermore, an Enhanced Occlusion Attention Module (EOAM) is incorporated to strengthen the representation of occluded areas, suppress background interference, and reinforce spatial-channel attention, thereby improving the detection of small and partially occluded defects. Experimental results demonstrate that the proposed LDSE-YOLO model achieves superior overall detection performance on the NEU-DET and GC10-DET benchmark datasets, with mAP@0.5 improvements of 4.3% and 2.1%, respectively, compared to mainstream baseline models.

This study investigates the efficiency of numerous phishing detection techniques, concentrating on the proposed Convolutional Neural Network-based model that demonstrates superior performance over traditional and advanced methods. Attaining an exceptional 99.3% of accuracy, 99.93% of precision, 98.70% of recall, and 99.31% of F1 score, the typical leverages advanced methodologies such as the Whale Optimization Algorithm for feature selection and Zero Trust Verification to enhance robustness and efficiency. By detecting subtle patterns and differences, the CNN-based model outperforms traditional approaches like Decision Trees and KNN, as well as advanced models such as LSTM, making it a really reliable resolution for phishing detection. The model is further strengthened by its ability to integrate comprehensive feature analysis, adaptive learning, and dynamic classification capabilities. Proposed future enhancements include expanding datasets to encompass multilingual and diverse samples, incorporating hybrid models like transformers for improved feature extraction and optimizing the system for actual detection and deployment. These developments aim to recover the model’s scalability and applicability evolving cybersecurity environments, offering a comprehensive and reliable solution to effectively mitigate phishing threats in real-world scenarios.

Phishing attacks have developed very complex making the use of cultured detection technologies significant for the protection of sensitive information. The research presents a new framework for detecting phishing founded on the convolutional LSTM (ConvLSTM) network and neural ordinary differential equations (Neural ODEs) thereby refining detection robustness and accuracy. It joins a number of advanced methods like Gaussian noise injection for enhancing the simplification ability of the model and a hybrid Particle Swarm Optimization (PSO) and Sailfish Optimization (SFO) method for effective feature selection. In addition, it confirms safe decisions and regularly reviewing phishing risks by means of a genetic algorithm to improve the Zero Trust Implementation Framework. The optional model performs very well with F1 scores above 99 and recall, accuracy, and precision above 99. Associated to existing methods the proposed model performs significantly better reducing false positives and false negatives and contribute a higher degree of phishing detection accuracy. Future developments to the system will contain multi-modal data integration actual flexibility intelligible AI, and more scalability to handle important applications. Increasing the dataset to include changing phishing techniques could further improve detection services and support the system’s resistance to new threats.

Cloud virtual machines (VMs), which offer dynamic resource distribution and economical solutions, are extensively used for scalable and effective computing in a range of sectors. Cloud environments are getting more and more prevailing, but this has made them appealing prey for ransomware attacks, which can encrypt substantial data, interrupt services, and negatively impact functioning continuity. This study proposes a scheme for ransomware detection (RD) in cloud environments bedded on Swish-Activated Temporal Convolutional Networks (SA-TCNs) to prevail these difficulties. Utilising sophisticated temporal modelling with Swish-TCN and optimal feature selection with the Artificial Bee Colony (ABC) algorithm, the framework excels current methods like Pulse, SINN-RD, T-BECA, and BSFR-SH, achieving a high accuracy of 99.51%. The system enhances resources with low CPU and memory utilization per VM. It efficaciously identifies ransomware with rapid detection times, maintaining real time applicability and scalability across small to medium-scale cloud systems. This study provides a reliable and effectual way to mitigate ransomware risks in dynamic cloud systems while maintaining data security and service continuity.

This paper presents a Novel GAN-Based Privacy-Enhanced Intrusion Detection System (IDS) to address the challenges posed by IoT and IIoT technologies, which increase vulnerability to cyberattacks. Conventional IDS face issues like data imbalance, privacy concerns, and adapting to evolving attack patterns. The proposed model integrates Bidirectional Gated Recurrent Units (Bi-GRUs) with Generative Adversarial Networks (GANs). GANs generate synthetic data to mitigate data imbalance, enhancing the model's generalization capabilities, while Bi-GRUs ensure accurate classification of complex temporal attack patterns. The model achieves an impressive 99.96% accuracy, with high precision (99.34%), recall (99.54%), and F1-score (99.21%), outperforming existing methods like CNN+LSTM, RNN, and XG-Boost. The confusion matrix shows perfect classification for “DoS” and “Mirai” attacks, with minimal misclassifications between similar benign traffic types. The ROC curve's AUC of 0.99 and the closely aligned accuracy curves for training and validation further highlight the model's robustness. The study emphasizes the importance of GAN-based augmentation in balancing hostile and benign traffic classes and significantly improving classification accuracy for rare attack types. This approach addresses critical issues in data imbalance, privacy, and attack classification, marking significant progress in IDS development.

Non-intrusive load monitoring (NILM) improves energy efficiency and mitigates the greenhouse effect by effectively monitoring energy consumption. However, many methods fail to fully utilize raw signal information and require large amounts of labeled data for training. In this paper, we propose a novel method for NILM. First, raw load data are transformed into two views using strong and weak augmentations, with data locality enhanced through patching to capture more comprehensive load features. Next, an unsupervised contrastive learning framework is designed based on patch and context contrast, aiming to maximize similarity between different views of the same sample while minimizing similarity between different samples. This framework enables the encoder to learn effective feature representations from unlabeled load samples. Then, the learned representations are used by a classifier for load identification, relying on a small amount of labeled data for supervised training. Ultimately, the paper achieves a semi-supervised model that combines a self-supervised encoder with a classifier guided by partial label information to enhance NILM performance. Experiments on PLAID, WHITED and REDD public datasets demonstrate the effectiveness of the proposed algorithm, which outperforms comparative algorithms in performance.

The aim of personalized education is to tailor the learning experience in accordance with individual students’ academic needs and maximize their academic performance through these customized approaches. However, the unavailability of real-time insights and personalized guidance poses a major challenge in the prediction of students’ grades and the development of customized study plans. To overcome these limitations, our research proposes the Student Performance Analysis Tool (SPAT), a machine learning-based analytical platform implemented in Python. SPAT was built using a dataset of 145 university students from U.S. university data. The tool focuses on predicting student performance based on academic and non-academic features which is novel in execution. Further, the tool investigates on seven machine learning models such as Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), Gradient Boosting, XGBoost, and AdaBoost. Also, SPAT combines real-time data visualization and grade prediction with an interactive interface, allowing teachers and students to make data-driven, informed decisions for academic enhancement. Results represented by SPAT indicates high accuracy for predicting the student rate for algorithms Gradient Boost and XGBoost as 79%. Through the use of AI-powered analytics, the tool proposes adaptive learning methods, showing the power of machine learning in developing data-driven with tailored education.

Tunnel construction relies on large shield machines for excavation. During shield tunneling, the effective operation of the sealing system at the tail of the shield machine is crucial. To prevent sealing failure, maintaining the pressure balance inside and outside the sealing cavity is essential. Traditional methods indirectly assess the grease pressure inside the cavity by predicting grease consumption, but they are prone to errors and uncertainties. Nowadays, with advancements in sensor technology, it is possible to directly measure grease pressure, making direct prediction feasible. However, because the causal relationships between covariates such as shield tail attitude, slurry pressure, grease injection pressure and grease pressure are directional and sparse. Traditional time series forecasting techniques struggle to capture these causal relationships, which reduces the model performance. To address the above-mentioned issues, we propose a causal contrastive transformer(CCformer) model for the first time to directly predict the grease pressure. CCformer accurately models the causal relationships of covariates on the grease pressure through the dynamic influencer. Meanwhile, it makes good use of contrastive learning to gather the distributed similar causal relationships together, demonstrating a memory ability for them. The experimental results prove that CCformer performs better than other state-of-the-art models on the real shield tunneling construction dataset, with the average absolute error reduced by 6.6% and the mean squared error reduced by 5.6%. This research provides reliable support for ensuring the balance of pressure inside and outside the sealing cavity and contributes to improving construction safety.