Concurrency and Computation-Practice & Experience最新文献

Multimodal sentiment analysis (MSA) has recently encountered two major challenges: non-textual modalities are often affected by noise, and sentiment intensity differences are difficult to capture. To address these issues, we propose a Sentiment Intensity Contrastive Text-Enhanced Fusion Network (SICTEF Net), which achieves deep collaboration among text, audio, and visual modalities through three key mechanisms. First, a grouped-channel-attention based Feature Enhancement Module (EMA) is designed to mitigate modality-specific noise and emphasize emotion-sensitive cues by combining spatial–channel interaction mapping with dual-branch attention fusion. Second, a text-centered cross-modal fusion mechanism is introduced, where bidirectional multi-head self-attention and a residual-enhanced encoder jointly enable complementary mappings between text and non-text modalities, thereby producing intermediate representations that preserve semantic primacy while incorporating fine-grained complementary information. Third, a sentiment-intensity weighted contrastive learning strategy dynamically assigns weights to positive and negative sample pairs according to their sentiment intensity differences, allowing the model to more precisely distinguish samples with varying degrees of similarity in the embedding space. Experimental evaluation on the CMU-MOSI and CMU-MOSEI datasets demonstrates that SICTEF Net consistently outperforms state-of-the-art baselines in binary accuracy, F1 score, seven-class accuracy, mean absolute error (MAE), and Pearson correlation. Comprehensive ablation studies further confirm the complementary benefits of EMA, the text-enhanced Transformer, and sentiment-intensity contrastive learning. These results indicate that combining text-driven deep interaction, non-text modality enhancement via channel attention, and contrastive learning can improve the accuracy and robustness of multimodal sentiment analysis.

Mobile application recommendation has emerged as a pivotal domain within the realm of personalized recommendation systems. Traditional mobile application sequence recommendation approaches are predominantly dedicated to the pursuit of sophisticated sequence encoders to achieve more precise representations. However, existing sequence recommendation methods primarily consider the sequential order of historical App interactions, overlooking the time intervals between applications. This oversight hinders the model's capability to fully unearth the temporal correlations in user behavior, consequently limiting the accuracy and personalization of mobile application recommendations. Moreover, the interactions between users and mobile applications are typically sparse, which weakens the model's generalization capabilities. To address these issues, we propose a novel method for mobile application sequence recommendation, incorporating time interval-aware attention and contrastive learning (called Ti-CoRe). Specifically, this approach introduces a novel sequence augmentation strategy based on similarity replacement within a contrastive learning framework. By considering textual similarities between applications, this method selectively replaces applications that possess lower similarity scores to generate augmented sequences, increasing the diversity of the sample space and mitigating data sparsity. Furthermore, integrating a time interval-aware mechanism into the BERT4Rec model, the paper presents a new T-BERT encoder. It precisely assesses the influence of fluctuating time intervals on the prediction of the subsequent mobile application, thereby ensuring a more nuanced app representation. Experiments conducted on the 360APP real dataset demonstrate that Ti-CoRe consistently outperforms various baseline models in terms of NDCG and HR metrics.

One of the life-threatening health conditions that affects millions of people across the globe is chronic kidney disease. Early detection and classification play an important role in treating and controlling the disease progression. The traditional healthcare system faces various challenges, like the risks such as progression of end-stage renal disease, high morbidity and mortality due to the increasing number of chronic kidney disease patients across the world. Millions of lives can be saved with early detection and proper treatment. This study proposes a novel optimized explainable cross attention transformer-based separable neural network model to detect and classify chronic kidney disease. In this model, the preprocessing approaches like image resizing, data augmentation, image normalization, missing data handling, data encoding, and data imputation are used to clean the data. Then, to choose the optimal attributes from the preprocessed data, a separable convolutional network and a transformer encoder are utilized. The softmax activation function and the fully connected layers in the classification layers perform the multiclass classification of this data. The interpretability and transparency of the model are improved using local interpretable model-agnostic explanations, and the convergence rate and training are enhanced with the integration of stochastic gradient descent optimizer. Two publicly accessible kindney disease-related datasets are used in this work to validate the model performance. The experiments conducted indicated that the proposed model attained the superior accuracy value of 98.45% and the lowest error rate of 1.55% and showed its superiority over the existing techniques.

Data deduplication plays an important role in modern data management as it reduces storage costs and ensures consistency by eliminating redundant records. The traditional data deduplication methods are effective for exact matches but struggle with adaptability and detecting near-exact duplicate records in unstructured or complex data. Machine learning (ML) addresses these limitations by using pattern recognition, feature learning, and statistical modeling to identify subtle similarities between records. This review classifies ML-based deduplication techniques into supervised, unsupervised, semi-supervised, and deep learning methodologies. It also discusses key challenges, including class imbalance, model interpretability, and computational overhead. The paper also explores recent developments in federated learning, real-time deduplication, and multimodal techniques to highlight current trends in these areas. Finally, the paper identifies key open issues and proposes a unified perspective for scalable, real-time deduplication systems that can accommodate diverse data types, structures, and system requirements.

The rapid adoption of electric vehicles (EVs) highlights the need for intelligent systems to improve energy efficiency and optimize driving range. Since energy consumption and driving range modeling are closely related, understanding the energy consumption (EC) of EVs can provide essential insights to drivers and reduce “range anxiety.” Previous studies have relied on traditional analytical and statistical methods, which lack the representativeness of influential factors and the interpretability of the model applied in EC modeling. To address this issue, we propose an explainable ensemble machine learning model to predict EC of EVs, considering the most important features and the factors that exhibit greater influence on EC. The Spritmonitor public real-world dataset is used for this study. First, data preprocessing is conducted before feeding data into the ensemble method. Second, the Energy Consumption Rate (ECR) was predicted using Gradient Boosting Regression Trees (GBRT). The proposed predictive framework demonstrates superior prediction accuracy compared to baseline models. GBRT achieved the highest R ² (1 and 0.99 for training and testing, respectively) and the lowest MAE (0.08) and RMSE (0.16) compared to other models, including XGBoost, LightGBM, and CatBoost. Finally, SHAP (Shapley Additive exPlanations) analysis was applied to explain the proposed model and identify the most influential dynamics factors, including driving range, capacity, speed, state of charge (SOC), ambient temperature, road type, driving style, air conditioning, and heating usage. The results suggest that the proposed framework can effectively enhance the prediction of the EC of EVs and facilitates the analyze driving factors, thereby supporting intelligent trip planning, adaptive energy-aware management in transportation systems and provide insightful feedback to drivers.

Accurate forecasting of carbon emissions from power generation enterprises is essential under China's dual-control policy. Although deep learning methods show strong potential, studies on their optimal configuration remain limited. This paper proposed a hybrid deep learning framework integrating a convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and an attention mechanism for carbon emission prediction in natural gas power plants. The present study utilized two distinct optimization methodologies: a structured design strategy encompassing light, medium, and heavy configurations, while the other employed Bayesian optimization for hyperparameter tuning. The models were evaluated using 5-fold cross-validation on 619 operational samples from two 487.1-MW condensing units in a power plant in Hainan, China. The medium configuration achieved the best balance between accuracy, efficiency, and stability, with R² = 0.9833, RMSE = 0.0342, and MAE = 0.0242. Under small-sample conditions, the structured design approach outperformed Bayesian optimization by 0.16% in accuracy while requiring only 7.42% of the training time. The proposed framework provides an efficient and interpretable reference for selecting deep learning architectures in small-sample industrial regression tasks and supports intelligent, low-carbon power generation applications.

Library inventory is vital for collection management and reader satisfaction. Conventional manual methods cannot support real-time updates, while existing automated solutions relying on centralized cloud computing suffer from bandwidth and latency limitations. To address these issues, we propose an edge-cloud collaborative real-time book inventory system. Spine detection and text recognition are executed on embedded edge devices, while the cloud handles rapid data retrieval to balance timeliness and accuracy. We design lightweight models for edge deployment, including the Library You Only Look Once (Lib-YOLO) detector with a StarNet backbone, shared convolutional head, and dual-scale hierarchical detection, supporting rotated objects for robust spine extraction. The optimized paddle practical optical character recognition (PP-OCR) pipeline removes text rectification and integrates a filtering algorithm to reduce redundant computation and improve efficiency. Deployed on an NVIDIA Jetson Nano, the system achieves 73 ms spine detection latency, 191 ms text recognition latency, and 97.1% overall accuracy under simulated library conditions. The Lib-YOLO model contains only 1.39 M parameters with 99% mean average precision (mAP), demonstrating the feasibility of precise, real-time inventorying in resource-constrained embedded environments.

Privacy-Preserving Data Sharing (PPDS) masks the individual's collected data (e.g., medical healthcare data) before being disseminated by organizations for analysis and research. Patient data contains sensitive values that must be dealt with while ensuring certain privacy conditions are met. This minimizes the risk of re-identification of an individual record from the group of privacy-preserved data. However, with the advancement in technology (i.e., Big Data, the Internet of Things (IoT), and Blockchain), the existing classical privacy-preserving techniques are becoming obsolete. In this paper, we propose a blockchain-based secure data sharing technique named “SecureChain”, which preserves the privacy of an individual record using local differential privacy (LDP). The three distinguished features of the proposed approach are lower latency, higher throughput, and improved privacy. The proposed model outperforms the benchmarks in terms of both latency and throughput. In terms of precision, the proposed method improves the accuracy to 88.53% compared to its counterparts, which achieved 49% and 85% accuracy. The results of the experiment verify that the proposed approach outperforms its counterparts.

For cross-city traffic prediction, the significant heterogeneity of traffic data across cities and the requirement for privacy protection make it challenging for conventional centralized spatiotemporal graph modeling techniques to balance predictive performance and data security. Therefore, this paper proposes AT-SPNet, a personalized federated spatiotemporal modeling approach specifically designed for cross-city traffic prediction. This method decouples the spatiotemporal modeling paths through the construction of a shared temporal branch and a hidden local spatial branch, thereby mitigating the heterogeneity of cross-city traffic data while preserving privacy. In the temporal branch, Gated Recurrent Units and a multi-head attention mechanism are incorporated to capture temporal dependencies, and a Squeeze-and-Excitation module is employed to enhance the extraction of informative features. In the spatial branch, a Spatial Attention Fusion module based on a triple-attention mechanism is designed to capture spatial features from multiple spatial perspectives, combined with static graph convolution and dynamic graph attention to construct a dual-modal information fusion path. Furthermore, to alleviate the adverse effects of cross-city data heterogeneity in federated training, a personalized federated learning strategy is introduced, which enables differentiated fusion of client spatial features without sharing raw data. Experiments on four real-world traffic datasets demonstrate that AT-SPNet outperforms existing methods in both prediction accuracy and cross-city generalization, validating the effectiveness and practical applicability of the proposed approach for cross-city traffic prediction.