The increasing volume, diversity, and complexity of Big Data require clustering techniques capable of handling high-dimensional, noisy, and heterogeneous datasets while maintaining scalability and robustness. This study proposes a BZ-optimized [Formula: see text]-medoids clustering algorithm (BZ-[Formula: see text]Medoids) that integrates fuzzy membership modeling, BZ-based medoid optimization, and distributed processing within a unified clustering framework. Fuzzy membership enables partial assignment of data points to multiple clusters, improving the representation of overlapping cluster boundaries and uncertain data regions. BZ optimization refines medoid selection using membership information, enhancing cluster compactness, stability, and resistance to outliers. Distributed execution further supports efficient large-scale processing through parallel computation across multiple nodes. The proposed framework addresses key limitations of conventional clustering approaches, including sensitivity to initialization, rigid cluster assignments, and reduced effectiveness in complex data environments. Experimental evaluation was conducted using the Bank Marketing dataset and compared with [Formula: see text]-means and Agglomerative clustering methods. The proposed approach achieved a clustering error of 0.38, substantially lower than the values of 3.55 and 3.59 obtained by [Formula: see text]-means and Agglomerative clustering, respectively. BZ-[Formula: see text]Medoids also achieved an accuracy of 0.99, outperforming [Formula: see text]-means (0.92) and Agglomerative clustering (0.91). Although the average convergence time was 12.35 s compared with 4.87 s for [Formula: see text]-means, the improved clustering quality, robustness, and scalability provide a favorable trade-off between efficiency and performance. The proposed framework achieved superior clustering quality while maintaining scalability for large and heterogeneous Big Data environments. These findings demonstrate the effectiveness of BZ-[Formula: see text]Medoids for adaptive, scalable, and high-quality clustering in practical analytics applications.
Early and accurate prediction of cardiovascular diseases (CVDs) remains challenging because existing diagnostic models often have limited generalizability, insufficient multimodal feature representation, and delayed detection. To address these limitations, a Pyramid Convolution Vision Transformer (PC-ViT)–Enhanced You Only Look Once version 10 (YOLOv10) framework is proposed to integrate clinical parameters with echocardiographic features for efficient multimodal heart disease prediction.The framework integrates pyramid convolutions to capture multi-scale cardiac structural patterns and Vision Transformer (ViT) layers to learn longrange dependencies across multimodal datasets. Additionally, YOLOv10 enables rapid and precise detection of defects in echocardiograms. The primary objective is to enhance prediction accuracy, minimize misclassifications, and offer explainable insights for clinical decision-making.Experimental evaluation demonstrates superior performance over existing hybrid CNN- and ViT-based models, yielding higher diagnostic accuracy, sensitivity, and F1 scores alongside fewer false alarms. The model's robust multimodal learning capability highlights PCViT-YOLOv10 as an efficient tool for early heart disease detection and timely intervention.
In the film industry, box office predictions are crucial for making investment decisions, marketing plans, and distribution choices. Current prediction approaches mainly use movie metadata or single-modal information, ignoring trailer video content and audience responses, resulting in low prediction accuracy and robustness. The proposed framework is novel in that it uses a Convolutional Neural Network (CNN)–Bidirectional Long Short-Term Memory (BiLSTM)–Attention (CNN–BiLSTM–Attention) regression and Artificial Rabbits Optimization (ARO)-based feature selection to integrate trailer visual features, trailer audio features, movie metadata, and audience responses into a single multimodal architecture. The proposed framework used the Movies YouTube Trailers & Sentiment dataset, which was preprocessed using mean value imputation, followed by Min–Max normalization and one-hot encoding. Vision Transformer (ViT) was used to extract visual characteristics from the trailer, and Mel-Frequency Cepstral Coefficients (MFCC), chroma, spectral centroid, zero-crossing rate, and Root Mean Square (RMS) energy were used to extract audio features. Audience response features were derived from sentiment scores, favorability, and vote counts. Visual, audio, metadata, and audience response features were fused into a unified multimodal representation. In order to forecast box office revenue, ARO chose the CNN–BiLSTM–Attention network's most informative feature subset. Mean Absolute Error (MAE) of 0.0246, Mean Squared Error (MSE) of 0.0009, Root Mean Squared Error (RMSE) of 0.0304, Mean Absolute Percentage Error (MAPE) of 0.0586, and Coefficient of Determination (R 2 ) of 0.9549 were all attained by the suggested model. The results showed that the proposed model improved prediction accuracy and provided an effective decisionsupport tool for movie producers and distributors.
The Boolean Satisfiability Problem is a central challenge in artificial intelligence and combinatorial optimization, with applications in circuit verification, planning, and bioinformatics. This study makes two main contributions: a systematic review of metaheuristic approaches conducted under the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) protocol, and a comparative evaluation of 11 algorithms. The considered methods include genetic algorithms, genetic algorithms with stepwise adaptation of weights, hybrid grey wolf genetic algorithms, genetic-based local search, population-based simulated annealing, dynamic clause weighting local search, stochastic local search, phase transition-based local search, random walk with local (1 + 1) evolutionary algorithm, improved adaptive genetic algorithm, and artificial bee colony. The evaluation relies on multiple criteria, including solution quality, robustness, success rate with confidence intervals, computational efficiency, and scalability across the benchmark families UF20 to UF250, complemented by additional benchmarks from the SAT Challenge Random Track (2018) and newly generated random 3-SAT instances. The results show that Genetic Algorithm, phase transition-based local search, dynamic clause weighting local search, genetic-based local search, random walk with local (1+1) Evolutionary Algorithm, and Artificial Bee Colony provide the most robust performance, maintaining a balance between exploration and exploitation. Beyond comparisons within metaheuristics, the study also examines how bio-inspired approaches compare with modern SAT solvers based on conflict-driven clause learning and stochastic local search across different problem scales. By contrast, highly specialized variants, such as genetic algorithms with stepwise adaptation of weights, population-based simulated annealing, hybrid grey wolf genetic algorithms, and improved adaptive genetic algorithms, degrade significantly on larger benchmarks. Overall, this study consolidates the strengths and weaknesses of metaheuristics and outlines future directions for adaptability, hybridization with machine learning, and scalable computations.
Recovering the three-dimensional trajectory of a shuttlecock from fixed monocular match videos is important for landing-point judgment, net-clearance analysis, flight-curve modeling, and stroke-quality assessment. Existing monocular badminton 3D recovery pipelines usually integrate court geometry, player pose, 2D shuttle trajectories, hit events, and physical constraints, demonstrating the feasibility of this task; however, in real matches, the shuttlecock is small and fast, and long 2D gaps, local large deviations, and hit-event misjudgments may still be amplified during 3D modeling, while frame-wise 3D ground truth is also scarce. To address these problems, this paper proposes a hitguided monocular 3D shuttlecock trajectory reconstruction method. The method first builds a 3Doriented robust 2D trajectory processing framework, improving the continuity and reliability of 2D inputs through multi-source candidate evidence, conservative selection, and candidate-level calibration; it then uses hit-to-hit flight segments as the basic unit and recovers segment-wise continuous 3D shuttlecock flight curves by combining court-camera geometry, 2D reprojection, pose semantics, and gravity-drag physical constraints. The evaluation is organized according to the task pipeline: 2D trajectory annotations are used to assess the quality of the preceding observation stage, ShuttleSet/ShuttleSet22 event annotations are used for posterior event-level consistency analysis, and curated high-confidence stereo3D clips are used for external 3D spatial reference validation. The results show that the proposed method increases F1@15px from 87.04% to 89.56% and reduces Danger@15 from 21.70% to 17.49% at the 2D observation level; under the given-hit protocol, the mean 3D error on stereo3D reference segments decreases from 2.680 m to 2.153 m, and the P90 3D error decreases from 8.291 m to 4.369 m. These results indicate that the proposed method can recover more continuous, stable, and physically interpretable 3D shuttlecock trajectories from real monocular match videos, providing a reliable spatial motion basis for stroke-quality assessment, tactical review, and sports video analysis.
