International Journal of Information and Communication Technology (19 papers in press)

Regular Issues

• Open Access

  A distributed and network-aware resource scheduling scheme for serverless cloud computing ( Free Full-text Access )

  by Yi Pan, Jiali You 
  Abstract: Serverless computing has become a mainstream cloud paradigm due to its elasticity in resource utilisation. However, it introduces high arrival rates and time-varying resource demands, making centralised network-aware scheduling approaches a bottleneck, especially in large resource pools. To address this, we propose a distributed, network-aware scheduling scheme where multiple agents make concurrent decisions. While this design improves scalability, it also brings challenges such as decentralised resource modelling, notification overhead, and decision contention. To mitigate these issues, we carefully design the DFaR placement algorithm and the MA migration algorithm. Simulations show that, compared to centralised approaches, DFaR achieves 23 orders of magnitude higher throughput with modest losses in scheduling objectives, and MA adapts to future resource fluctuations with performance comparable to prediction-based strategies.
  Keywords: serverless computing; distributed scheduling; scheduling speed; network awareness.
  DOI: 10.1504/IJICT.2025.10073516
   
  

• Open Access

  Siamese-based tennis movement gesture assessment via 3D tracking and spatio-temporal scoring ( Free Full-text Access )

  by Qing Miao, Chengzhao Li, Mingfang Wu 
  Abstract: This study proposes SiamAttn-3D + spatio-temporal scoring module (ST-ScoreNet), an end-to-end framework for objective tennis movement assessment. The SiamAttn-3D tracker employs 3D spatio-temporal attention to achieve robust joint localisation (84.6% success rate at >160 km/h racket speeds), overcoming motion blur and occlusion challenges. Joint trajectories feed into ST-ScoreNet, which integrates graph convolutions and bidirectional gated recurrent unit (GRUs) to model biomechanical constraints and temporal dynamics. Evaluated on the Tennis-ITF dataset, the system attains a 92.3% F1-score in stroke assessment (= 0.89 vs. coach ratings) a 6.9% improvement over state-of-the-art methods. Real-time processing at 23 frames per second (FPS) enables instantaneous feedback, reducing hardware costs by 83% compared to sensor-based solutions. Limitations include sensitivity to weather degradation and athlete anthropometrics, with federated learning proposed for future personalisation.
  Keywords: twin networks; posture evaluation; tennis motion analysis; spatio-temporal modelling.
  DOI: 10.1504/IJICT.2025.10073532
   
  

• Open Access

  Cross-border trade export prediction based on reinforcement learning and multimodal data ( Free Full-text Access )

  by Yang Yu, Ziwen Sun, Jiachen Li 
  Abstract: Cross-border trade export forecasting is important for enterprises to optimise resource allocation. However, existing prediction methods have the problem of insufficient single modal feature extraction, for this reason, this paper first optimises the reinforcement learning (RL) algorithm based on multilevel strategy and multilevel reward (MSRL). Then CNN, Doc2Vec model, and improved ResNet152 model were used to extract static variable features, comment text features, and image features of cross-border trade export sales volume, respectively, and a hierarchical attention mechanism was designed to fuse multimodal features. The hyperparameters of the BiGRU model are optimised using MSRL (MSRL-BiGRU), and the fusion features are input into MSRL-BiGRU, which efficiently and automatically searches for the optimal strategy and reduces the prediction error. The experimental results show that the proposed method improves the coefficient of determination R2 by 4.8418.67%, which can realise the accurate prediction of cross-border trade export sales.
  Keywords: cross-border trade export forecasting; reinforcement learning; multimodal fusion; hierarchical attention mechanism; BiGRU model.
  DOI: 10.1504/IJICT.2025.10073651
   
  

• Open Access

  A data compliance sharing algorithm for intelligent connected vehicles empowered by federated learning ( Free Full-text Access )

  by Bei Zhou 
  Abstract: With the rapid development of intelligent connected vehicle (ICV) technology, massive amounts of vehicle data have become an important resource for advancing intelligent mobility and autonomous driving technologies. However, the sharing of these data involves significant privacy leakage risks and compliance challenges, especially in multi-party collaboration scenarios. To this end, this paper proposes a federated learning (FL) framework integrating differential privacy and Paillier homomorphic encryption for intelligent connected vehicle (ICV) data sharing. The architecture integrates differential privacy (DP) and homomorphic encryption (HE) through four functional layers. The four-layer architecture achieves 93.5% model accuracy with 79% lower privacy leakage risk (0.21 vs. 0.98 baseline) on 50,000 driving scenarios. Momentum-accelerated averaging and 8-bit gradient quantisation reduce bandwidth consumption by 62%. Experimental validation demonstrates superior privacy-utility balance compared to standalone DP (0.6 risk) and HE (0.5 risk) implementations.
  Keywords: intelligent connected vehicle; ICV; federated learning; FL; data compliance sharing; privacy protection.
  DOI: 10.1504/IJICT.2025.10073652
   
  

• Open Access

  Web public opinion prediction of public events based on graph convolutional neural networks in big data ( Free Full-text Access )

  by Xianli Zeng, Hui Deng, Qizhong Luo, Lina Xiong 
  Abstract: Aiming at the problems of complex topology and strong temporal dynamics in the prediction of online public opinion on public events in big data environment, this paper proposes the spatio-temporal graph convolutional public opinion prediction model that fuses heterogeneous graph convolution and time domain convolution. The framework employs spectral graph convolution for topology modelling and dilated temporal convolution for dynamic dependency capturing. A multi-entity graph is constructed based on public health emergency management ecosystem and microblog rumour database, and feature fusion is achieved through cross-platform attention mechanism. Experiments show that the model has an F1 value of 89.2% in public opinion detection and a heat prediction RMSE of 6.31, outperforming state-of-the-art baselines by 12.7% and 31.5% respectively. It can warn 93% of high-risk events 52 minutes in advance, enabling proactive intervention for public governance.
  Keywords: graph convolutional neural network; opinion prediction; heterogeneous graph neural network; spatio-temporal modelling.
  DOI: 10.1504/IJICT.2025.10073653
   
  

• Open Access

  Data element perspective: green credit risk assessment using a multi-layer deep neural network ( Free Full-text Access )

  by Lei Gu, Tao Liu 
  Abstract: Against the backdrop of accelerating double carbon goals, green credit has become a key tool for channelling funds into low-carbon sectors, requiring sophisticated risk assessment models. Conventional approaches, limited by single-dimensional data and poor dynamic adaptability, fail to address green projects multi-faceted risks (long investment cycles, rapid technological changes, strong policy dependency). This study proposes a novel green credit risk assessment framework from a data element perspective, using a multi-layer deep neural network (MLDNN). It integrates multi-source heterogeneous data, employs a three-tier neural architecture with an attention mechanism, and uses an adaptive learning rate algorithm. Empirical results from a provincial bank show the model achieves 92.57% risk identification accuracy, 11.2% higher than traditional BP neural networks, with notably improved generalisation in small-sample scenarios.
  Keywords: green credit; risk assessment; data elements; multi-layer deep neural network; MLDNN; attention mechanism.
  DOI: 10.1504/IJICT.2025.10073679
   
  

• Open Access

  Development of an AI-powered classification model for monitoring construction disturbances using remote sensing data ( Free Full-text Access )

  by Jing Yan, Yin Liu, Shengcheng Xie, Yimin An, Yifan Liang 
  Abstract: In complex settings, the identification and dynamic monitoring of construction disturbance areas still face problems such as insufficient feature extraction, limited generalisation, and unstable multi-temporal detection accuracy. This study proposed a novel multi-level integrated approach that combines fractal network evolution algorithm (FNEA) segmentation, genetic algorithm (GA) global optimisation, and convolutional neural network (CNN) multi-scale feature learning to achieve high-precision disturbance recognition and dynamic monitoring. Experimental results showed that the method achieved an overall accuracy of 95.2% +-0.4% (95% CI [95.0, 95.4]), maintained an accuracy above 99% in multi-temporal tests, reduced the false alarms and missed detections by 0.7-5.2% compared with baseline methods, and converged within 30 iterations. Compared with existing techniques, the framework provides an intelligent and efficient solution through the joint use of image segmentation, evolutionary optimisation, and deep feature learning, opens a new direction for remote sensing monitoring in complex construction environments.
  Keywords: construction disturbance detection; fractal network evolution; genetic algorithm; convolutional neural network; CNN; multi-feature fusion.
  DOI: 10.1504/IJICT.2025.10073710
   
  

• Open Access

  A news content recommendation model integrating social relationships and temporal features ( Free Full-text Access )

  by Xiaofeng Liang 
  Abstract: Aiming at the problems of dynamic changes of user interests and underutilisation of social influence in current news recommendation, this paper proposes a social-temporal enhanced news recommendation model (STENR) that integrates social relations and temporal features. The model uses graph neural network (GNN) to model the user-user social relationship graph, and adopts transformer encoder to capture temporal dependencies and generate temporal embeddings reflecting the dynamics of users recent interests. At the same time, a text encoder is used to extract the deep semantic features of the news content. The users comprehensive interest representation is generated dynamically by weighting the fusion information adaptively through the attention mechanism. Experiments show that the AUC of STENR is increased to 0.812, the length of user stay is increased by 23.4%, and the social conversion rate is increased by 15.3%, which verifies its academic validity and industrial value.
  Keywords: news recommendation; social relationship modelling; temporal feature extraction; graph neural network; GNN.
  DOI: 10.1504/IJICT.2025.10073711
   
  

• Open Access

  Analysis of English sentiment semantic evolution based on BERT and dynamic word embeddings ( Free Full-text Access )

  by Yueqin Liu 
  Abstract: To address the challenge of accurately capturing the evolutionary trends of English sentiment semantics in large-scale time-series text data, this study proposes a sentiment semantic evolution analysis method by fusing BERT and dynamic word embeddings. First, the overall framework of the fusion model is constructed, including the data preprocessing layer, feature extraction layer, fusion layer, and application layer. Second, based on the theory of semantic change, the sentiment semantic evolution analysis index system is determined, covering temporal stability, contextual similarity, and sentiment polarity variation. Key features of sentiment semantics are extracted from time-slice corpora. Application results on a historical English corpus show that the model’s semantic evolution prediction accuracy reaches 89.72%, and the time efficiency is improved by 15.3% compared with single models, demonstrating excellent performance in capturing temporal dynamics and semantic accuracy.
  Keywords: English sentiment semantics; semantic evolution; BERT; dynamic word embeddings; feature fusion.
  DOI: 10.1504/IJICT.2025.10073712
   
  

• Open Access

  Social media sentiment diffusion modelling algorithms for brand communication ( Free Full-text Access )

  by Juntao Xie 
  Abstract: Addressing the challenge of accurately predicting and guiding the spread of brand sentiment on social media, this study proposes an analytical framework that integrates dynamic communication modelling with intelligent algorithms. A two-layer coupled communication model (DI-SCIR) is constructed to quantify the migration patterns of sentiment by integrating users time-varying behaviour and cross-platform interaction mechanisms. A three-dimensional influence strength measurement method (WSD-Rank) is designed to identify key communication nodes based on coverage, timeliness, and forwarding depth. Combines AI clustering algorithms to uncover primary diffusion pathways and develops generative intervention strategies to achieve sentiment guidance and risk warning. Empirical verification shows that this method achieves an accuracy rate of 89.2% in brand sentiment prediction, providing effective theoretical modelling and algorithmic support for risk management and strategy optimisation in brand communication.
  Keywords: brand communication; social media sentiment analysis; AI algorithms; sentiment prediction.
  DOI: 10.1504/IJICT.2025.10073713
   
  

• Open Access

  An intelligent analysis model for enhancing rural e-commerce sales efficiency in live streaming environments ( Free Full-text Access )

  by Rong Zhou 
  Abstract: Amidst the swift advancement of internet technology, live broadcast e-commerce has emerged as a pivotal driver for rural economic growth. However, the issue of low sales efficiency is becoming increasingly conspicuous, necessitating the development of effective solutions. Against this backdrop, this study investigates strategies to enhance the sales efficiency of rural e-commerce within the live broadcast context, with a focus on devising an intelligent analysis model to address the sales challenges. By precisely analysing the vast amounts of multi-source data generated during live broadcasts, the model employs data mining, machine learning, and deep learning algorithms for in-depth association analysis and feature extraction. Experimental results demonstrate that compared with outstanding deep learning models, the REISE model has reduced errors by approximately 25.60% and 22.31% in MAE metrics, 49.46% and 46.77% in MSE metrics, and 30.12% and 29.27% in RMSE metrics, respectively.
  Keywords: rural e-commerce; live broadcast sales; intelligent analysis model; sales efficiency; data mining.
  DOI: 10.1504/IJICT.2025.10073714
   
  

• Open Access

  Financial risk prediction and warning system based on SGAN deep learning ( Free Full-text Access )

  by Linjing Dang 
  Abstract: In response to the problem of insufficient performance of traditional financial risk warning models in class imbalanced data, this study proposes a deep learning warning method based on semi supervised generative adversarial network (SGAN). Firstly, construct a financial feature system and perform data balancing processing through SMOTE Tomek mixed sampling. The model adopts a generator discriminator dual network architecture and constructs a composite loss function using cross entropy loss and Wasserstein distance. The experimental section selects financial data of Chinese a-share listed companies from 2016 to 2020, and uses transfer learning strategy to fine tune the pre trained model in the manufacturing industry to the retail industry. The empirical results show that this method improves the F1 score (0.87) and AUC value (0.92) compared to traditional logistic regression, effectively solving the early warning problem caused by the temporal correlation and industry heterogeneity of financial data.
  Keywords: supervised generative adversarial network; SGAN; financial risk warning; SMOTE Tomek mixed sampling; generator discriminator dual network.
  DOI: 10.1504/IJICT.2025.10073820
   
  

• Open Access

  Application of CNN image signal denoising technology in virtual reality environment ( Free Full-text Access )

  by Hairong Wang, Lingling Wang 
  Abstract: The virtual reality environment mainly comprises panoramic images, providing an immersive visual experience. However, factors like shooting environment, lighting conditions, and image compression often lead to noises, affecting the image quality and the users immersive experience in the virtual environment. This paper implements a deep learning framework combining ResNet-50 and U-Net to effectively remove noise from panoramic images and improve the users immersive experience. ResNet extracts deep features of images through a residual learning mechanism, enhances the precision of image alignment, and reduces the possibility of noise expansion. U-Net adopts an encoding-decoding structure, which preserves image details and denoises through skip connections, avoiding over-smoothing, to improve the denoising effect. The results show that at different noise intensities, the method in this paper is significantly better than the mean filtering and Frost-filter methods in terms of peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). The PSNR value of the method in this paper is 36.89 dB when the noise intensity = 0.2, which is higher than that of the frost-filter (34.25 dB) and mean filtering (33.34 dB). Its SSIM values are above 0.90 under different noise intensities, which are higher than those of other methods. It can be found that the denoising model of U-Net + ResNet can comprehensively process various types of noise and show a better balance in restoring image details and structures, providing an effective solution for panoramic image denoising.
  Keywords: image denoising; convolutional neural network; panoramic image; residual network; virtual reality.
  DOI: 10.1504/IJICT.2025.10073821
   
  

• Open Access

  Mask-embedded transformer for English text recognition and correction ( Free Full-text Access )

  by Haiying Sang 
  Abstract: As the digital age moves quickly, automatic recognition and correction of English text has become a significant job in the field of natural language processing (NLP). Most traditional ways of correcting text use simple statistical models and manual procedures, which do not work well with complicated grammatical, spelling, and semantic mistakes. This paper suggests an English text recognition and correction framework called MT-Tec, which is based on the improved transformer model and the masked embedding technique. MT-Tec can find and fix spelling mistakes, grammar mistakes, and vocabulary mistakes through multilevel context modelling and accurate error correction mechanisms. The MT-Tec framework works very well with many kinds of text errors and text qualities, and it is especially good at handling low-quality text. In general, the MT-Tec framework can be quite helpful for automatic proofreading, revising text, and learning a new language.
  Keywords: English text recognition and correction; improved transformer; masked embedding; natural language processing; NLP.
  DOI: 10.1504/IJICT.2025.10073376
   
  

• Open Access

  Data-centric analytics for ideological sentiment monitoring: fusion of features with optimised attention mechanisms ( Free Full-text Access )

  by Wencong Wu, Juan Wang 
  Abstract: To address sentiment inaccuracies in civic education contexts, this study proposes a data-driven analytics framework integrating domain-adaptive feature engineering with hierarchical modelling. We construct Chinese social media corpora (Weibo/WeChat) through keyword-filtered crawling and interaction-weighted prioritisation, reducing noise by 42%. A hybrid feature space combines TF-IDF lexical patterns, syntactic POS distributions, Word2Vec/BERT embeddings, and HowNet-derived sentiment features. The core classification employs a Bi-LSTM model with attention mechanisms, dynamically weighting sentiment-bearing terms while compensating for category imbalance via class-weighted cross-entropy loss. Crucially, ideological semantics are mapped through logistic regression classifiers trained on annotated civic categories. Experimental results demonstrate: 1) attention weights effectively localise civic sentiment triggers; 2) domain feature fusion improves classification robustness; 3) semantic mapping achieves 89.2% accuracy in civic topic identification. This methodology enables real-time Kafka-based opinion monitoring while preserving interpretability for educational governance.
  Keywords: sentiment analysis; TF-IDF; social media monitoring; Bi-LSTM.
  DOI: 10.1504/IJICT.2025.10073437
   
  

• Open Access

  STIRS: cyber-physical trust integration for sustainable resource sharing in IoT-enabled cold chains ( Free Full-text Access )

  by Yuanzhao Zhao 
  Abstract: To address resource fragmentation and energy inefficiency in cold chain logistics, we propose sensor-trust integrated resource synergy (STIRS) - a blockchain-secured framework integrating internet of things (IoT) sensing with thermodynamic optimisation. The architecture establishes a dynamic trust model that quantifies hardware reliability metrics (battery decay, signal strength, calibration cycles) into adaptive credit weights via Sensor Health Index (SHI). Combined with mixed-integer programming and lightweight Byzantine consensus (0.48s latency), it enables real-time co-scheduling with: 1) energy consumption modelling incorporating ambient temperature sensitivity (dT/dt) and door-opening penalties (0.8-1.2 kWh/event); 2) General Data Protection Regulation (GDPR)-compliant homomorphic encryption. Validation using public United States Department of Agriculture - Agricultural Research Service (USDA-ARS) and commercial Taobao datasets (38,700 orders) demonstrates statistically significant improvements: 15-18% energy reduction, 23.7% resource utilisation gain, and 40% decrease in temperature deviation versus four benchmarks.
  Keywords: cold chain logistics; resource sharing model; IoT sensing technology; dynamic trust assessment; energy consumption optimisation.
  DOI: 10.1504/IJICT.2025.10073438
   
  

• Open Access

  AI-POA dual-engine framework: enhancing English speaking teaching through multimodal assessment ( Free Full-text Access )

  by Minmin Kong 
  Abstract: To address critical bottlenecks in production-oriented approach (POA) English speaking instruction - including high feedback delays, inefficient contextual task generation, and suboptimal resource allocation - this study proposes an AI-augmented POA framework. We developed a dual-engine architecture integrating dynamic task generation, multimodal resource recommendation, and multidimensional assessment to optimise POA's 'drive-facilitate-evaluate' closed loop. In a 12-week quasi-experiment with 120 computer science graduates, the experimental group (AI-POA) demonstrated significantly higher oral proficiency gains versus traditional POA controls (36.1% vs. 19.2%, p < 0.001), with content elaboration increasing by 22.6%. The framework reduced instructor feedback time per task from 8.2 to 0.3 minutes (27-fold improvement) and lowered cognitive load (NASA-TLX: 42 vs. 65, p < 0.001). Task acceptance reached 92% through cognitive-contextual difficulty adaptation. This work establishes an AI-POA synergy that enhances pedagogical outcomes while substantially alleviating instructor workload.
  Keywords: production-oriented approach; POA; AI collaboration; English speaking teaching; multimodal assessment; adaptive learning.
  DOI: 10.1504/IJICT.2025.10073439
   
  

• Open Access

  Causality mining for historical events based on knowledge graphs ( Free Full-text Access )

  by Xin Li 
  Abstract: This study proposes a novel probabilistic inference framework leveraging knowledge graphs (KG) to address sparsity and implicitness challenges in historical event causality. Key innovations include a dynamic event embedding (DEE) model incorporating a temporal decay factor β to capture the dynamic weakening of causal strength over time, and a causal graph neural network (CauGNN) utilising directional propagation and cross-event attention for modelling causal transmission between discontinuous events. Evaluated on the event-centric knowledge graph (EventKG) dataset spanning centuries, the method achieves 89.2% causal inference accuracy - a significant 12.7% improvement over state-of-the-art approaches - and a low temporal prediction deviation of 5.2 years. This work establishes a mathematical model for historical causal decay, shifts computational historiography toward quantitative causal reasoning, and provides verifiable tools for historical analysis, education (via the HistVis platform), and societal risk extrapolation.
  Keywords: causal inference; knowledge graph; dynamic event embedding; DEE; historical event analysis.
  DOI: 10.1504/IJICT.2025.10073440
   
  

• Open Access

  Cross-modal Chinese text representation enhancement for multimodal sentiment analysis ( Free Full-text Access )

  by Xin Zhang 
  Abstract: Addressing the dual challenges of textual vulnerability to noise and inefficient cross-modal interaction in Chinese multimodal sentiment analysis, this paper introduces a novel framework enhanced by a cross-modal text enhancement module (CTEM). The CTEM adaptively recalibrates semantic representations of Chinese text through contextual refinement. Concurrently, a cross-modal attention mechanism directs visual and acoustic feature extraction, enabling synergistic fusion across modalities. Evaluated on the Chinese single and multimodal sentiment (CH-SIMS) benchmark (featuring unaligned video segments and dual sentiment labels), our model achieves 83.2% accuracy - surpassing mainstream baselines by up to 3.2% with a 0.029 F1-score gain. Ablation studies confirm the critical contributions of both the CTEM representation refinement and cross-modal interaction design. This work establishes a robust paradigm for decoding nuanced sentiment in linguistically complex Chinese multimedia content.
  Keywords: cross-modal text information enhancement; multimodal sentiment analysis; Chinese semantic understanding; feature fusion; attention mechanism.
  DOI: 10.1504/IJICT.2025.10073441