Forthcoming Articles
International Journal of Information and Communication Technology
Forthcoming articles have been peer-reviewed and accepted for publication but are pending final changes, are not yet published and may not appear here in their final order of publication until they are assigned to issues. Therefore, the content conforms to our standards but the presentation (e.g. typesetting and proof-reading) is not necessarily up to the Inderscience standard. Additionally, titles, authors, abstracts and keywords may change before publication. Articles will not be published until the final proofs are validated by their authors.
Forthcoming articles must be purchased for the purposes of research, teaching and private study only. These articles can be cited using the expression "in press". For example: Smith, J. (in press). Article Title. Journal Title.
Articles marked with this shopping trolley icon are available for purchase - click on the icon to send an email request to purchase.
Online First articles are also listed here. Online First articles are fully citeable, complete with a DOI. They can be cited, read, and downloaded. Online First articles are published as Open Access (OA) articles to make the latest research available as early as possible.
Articles marked with this Open Access icon are Online First articles. They are freely available and openly accessible to all without any restriction except the ones stated in their respective CC licenses.Register for our alerting service, which notifies you by email when new issues are published online.
International Journal of Information and Communication Technology (63 papers in press) Regular Issues
Abstract: Faced with the challenge of clustering complex graph structures in big data, traditional methods often separate graph preprocessing and deep representation learning, leading to suboptimal cognitive-resource allocation. Inspired by cognitive load theory (CLT), this paper proposes a co-clustering model that combines adaptive graph transformation and structural self-attention. The model actively reduces the intrinsic cognitive load of the input graph structure through a learnable graph transformer and efficiently allocates the associated cognitive load to construct clusters through a self-attention mechanism that incorporates structural priors. Experiments on real datasets, such as Amazon-Computers, show that the accuracy of the proposed model reaches 0.852, which is significantly improved by 4.0% compared with the optimal baseline (p < 0.001), and ablation experiments confirm the substantial contribution of each module. Our research results offer a new paradigm for combining cognitive theory with algorithm design in graph learning. Keywords: graph clustering; cognitive load theory; CLT; adaptive graph transformation; structural self-attention; big data. DOI: 10.1504/IJICT.2026.10079460
Abstract: Consumer reviews often contain the anchoring effect, a cognitive bias where users express emotions based on a reference point such as price or a comparison object. Traditional sentiment classification models struggle to distinguish between the anchor point and the true attitude, leading to misjudgements. To address this, we propose a fine-grained sentiment classification method that integrates anchoring psychological bias. By jointly extracting numerical and semantic anchors, we explicitly encode cognitive bias as learnable feature variables. A dual-channel interaction network dynamically fuses semantic and bias features. Experiments on Amazon and Yelp datasets show accuracies of 89.2% and 86.7%, improving by 3.5% and 2.9% over the baseline bidirectional encoder representations from transformers model, with area under the curve increases of 4.1% and 3.3%. These results verify that incorporating anchoring psychological features improves complex review sentiment classification, offering an interpretable cognitive perspective on consumers irrational expressions. Keywords: fine-grained sentiment classification; anchoring effect; cognitive bias; consumer reviews. DOI: 10.1504/IJICT.2026.10079461
Abstract: Online teaching platforms generate rich interaction traces, yet real-time evaluation still struggles to capture how learner understanding shifts during instruction. To address this gap, this study proposes a dual-stream framework for real-time cognitive transfer tracking in online teaching evaluation. First, behavioural evidence and semantic-cognitive evidence are encoded separately to preserve temporal rhythm and instructional meaning. Then, a cross-stream interaction mechanism aligns the two evidence sources and feeds a confidence-aware tracking module that stabilises state updates under noisy streaming inputs. Finally, the tracked states are mapped to process-sensitive evaluation outputs. Experimental results show that the proposed method achieved macro-F1 scores of 0.792 and 0.776 on two datasets, reduced temporal variation by 21.2 % compared with the strongest baseline, and maintained stable step-wise inference at 22.4 milliseconds. The framework demonstrates strong accuracy, robustness, and practical real-time usability. Keywords: online teaching evaluation; cognitive transfer tracking; dual-stream network; real-time learning analytics. DOI: 10.1504/IJICT.2026.10079462
Abstract: The fragmentation and heterogeneity of student behaviour data hinder traditional methods from constructing complete digital profiles for early academic risk perception. To address semantic alignment issues in multi-source data and capture temporal behaviour evolution, this study proposes a data fusion framework integrating attention mechanisms and temporal deep learning. It combines data from campus cards, learning management systems, and psychological assessments to build student profiles. Experiments show the model achieves an area under the curve of 0.937 in risk prediction, 4.1% higher than extreme gradient boosting. In early perception tasks 8 weeks in advance, the normalised discounted cumulative gain at 5 reaches 0.884, identifying 85.7% of potential at-risk students. The research demonstrates that multi-source heterogeneous fusion and temporal modelling synergistically enable early precise intervention. Keywords: multi-source heterogeneous data; student digital profile; academic risk warning; time-series deep learning. DOI: 10.1504/IJICT.2026.10079479
Abstract: This paper proposes a blind source separation scheme based on an improved sine cosine algorithm (SCA) to address the challenge of separating multiple sound sources in music signals. By constructing a linear instantaneous hybrid model, introducing adaptive control parameters and local disturbance mechanisms to optimise the standard SCA, and constructing the objective function with reconstruction error and independence constraints, the optimal separation matrix is searched for. Experiments have shown that the improved SCA achieves an average SDR of 11.5 dB in two source scenarios, which is 1.55 dB higher than the standard SCA. In three source scenarios, the SDR is 10.2 dB, which outperforms traditional algorithms such as ICA and NMF. It has higher accuracy, stability, and noise resistance, providing an effective technical solution for blind source separation of music signals. Keywords: improved SCA algorithm; music signal; blind source separation; intelligent optimisation; signal processing. DOI: 10.1504/IJICT.2026.10079480
Abstract: To address the challenges of high noise and low latency in real-time data processing for wide-area digital metering devices in new energy grid connection environments, this paper proposes a lightweight spatio-temporal fusion network. This method synergistically integrates multi-head attention mechanisms with causal convolutions. A dual-channel noise filtering module is innovatively designed to dynamically suppress composite noise in complex electromagnetic environments, while an online incremental learning framework is incorporated to mitigate data distribution drift caused by fluctuating renewable generation. Validation tests on the Institute of Electrical and Electronics Engineers 37-node test system demonstrate that our method achieves a voltage/current measurement accuracy of 98.2% with an end-to-end latency below 15 ms. These results not only meet the stringent real-time monitoring requirements of smart grids but also provide a highly robust and adaptable solution for future wide-area power metering systems. Keywords: deep learning algorithms; wide area power metering; real-time data processing; online learning; noise suppression. DOI: 10.1504/IJICT.2026.10079481
Abstract: Relatively important nodes have become a key issue in the field of complex network research, and along with the advent of the era of big data, more and more novel techniques are involved in the research of complex networks. Different from the traditional method of analysing node importance based on network structure, this paper combines computer science methods with complex network significant node mining, and proposes a method of mining relatively important nodes based on label propagation SIGELAP (Significant node Identification algorithm based on Graph Embedding and LAbel Propagation algorithm). The method first maps the network into a vector representation using the network representation learning method. Then, the vector representation of the network is fed to the machine learning algorithm as input, and then label propagation algorithm is used to classify the nodes and to mine the relative nodes. Through experiments on real networks such as the 911 telephone network and the SARS virus propagation network, the SIGELAP algorithm proves to be accurate and applicable in mining relatively important nodes. Keywords: complex network; important nodes; label propagation; graph embedding. DOI: 10.1504/IJICT.2026.10079482
Abstract: With the rapid growth of fashion e-commerce and rising demand for personalised dressing, clothing compatibility calculation has become essential for intelligent fashion recommendations. To address the limitations of existing methods in capturing fine-grained visual features and modelling complex item-to-item matching relationships, this study proposes an end-to-end compatibility model combining transformer and graph convolutional networks (GCN). The model leverages transformer to extract fine-grained visual semantic features of clothing items, adopts GCN to model deep matching relationships between items, and outputs the final compatibility score through a learnable multi-branch weighted fusion module. Experimental results on the Polyvore and FashionVC datasets show classification accuracies of 92.35% and 91.78%, with a highest AUC of 95.12%, significantly outperforming five mainstream and two state-of-the-art (SOTA) models. In complex scenarios involving six items and cross-category combinations, accuracy remains above 89%, with over 93% consistency with human labelling. This robust intelligent matching solution can be widely applied in online retail, virtual dressing, and personalised fashion content production. Keywords: clothing matching compatibility; graph convolutional network; GCN; transformer; feature extraction. DOI: 10.1504/IJICT.2026.10079483
Abstract: Semantic communication overcomes traditional capacity limits by extracting the meaning of information; however, existing separative coding methods overlook differences in the semantic importance of images, resulting in critical semantic information being easily lost when channel conditions deteriorate. This paper proposes a joint source-channel coding scheme based on dual attention: spatial and channel attention are embedded at the encoding stage to dynamically focus on semantically critical regions, and the transmission dimension is adaptively adjusted according to the signal-to-noise ratio to achieve joint optimisation of semantics and the channel. Experiments on the CIFAR-10 and Kodak24 datasets demonstrate that at a signal-to-noise ratio of 10 decibel, the peak signal-to-noise ratio reaches 33.68 decibel, representing a 1.23 decibel improvement over the state-of-the-art method; under 0 decibel adverse conditions, it maintains 26.4 decibel, a 4.3 decibel improvement over traditional separative coding. This method effectively enhances both image transmission robustness and reconstruction quality. Keywords: semantic communication; attention mechanism; joint source-channel coding. DOI: 10.1504/IJICT.2026.10079484
Abstract: Performing arts instruction has long relied on subjective experience, making it difficult to achieve precise and personalised learning feedback. While traditional artificial intelligence methods can process quantitative data, they struggle to capture ambiguous concepts such as expressive movement and emotional delivery, posing a core challenge for assessment. To address this, this study innovatively integrates fuzzy logic with deep learning to construct an intelligent evaluation framework capable of understanding the grey areas of artistic expression. Experiments conducted on public dance datasets demonstrate that compared to traditional precise algorithms, this approach improves overall accuracy from 0.82 to 0.91 and significantly enhances the normalised discounted cumulative gain metric for ranking quality. This research validates the effectiveness of integrating fuzzy logic in reconstructing artistic teaching methods, offering a new pathway toward more human-centred and explainable intelligent artistic guidance. Keywords: fuzzy logic; performing arts education; personalised assessment; artificial intelligence; AI; multimodal. DOI: 10.1504/IJICT.2026.10079485
Abstract: The scale of resources in university libraries has grown exponentially. Traditional recommendation methods struggle to simultaneously model heterogeneous user-resource relationships and deep semantic features, leading to significant performance drops in cold-start scenarios. To address this, we propose dual-channel graph neural network with attention fusion, a dual-channel graph neural network with semantic fusion. It captures high-order topological structure via relation-aware graph convolution and extracts semantic representations from metadata using attention mechanisms. An adaptive gating mechanism dynamically fuses both features. On a dataset with 156,942 resources and 1.8 million borrowing records, the method achieves an area under the curve of 0.937 and NDCG@10 of 0.581, outperforming light graph convolution network by 3.2% and 4.5%. Generalisation is verified on the Amazon Books dataset. The dual-channel architecture effectively addresses data sparsity in academic resource recommendation, supporting smart library services. Keywords: resource recommendation; semantic fusion; adaptive gating; university library. DOI: 10.1504/IJICT.2026.10079486
Abstract: The development of the internet has driven the growth of e-commerce for agricultural products and has also advanced logistics. However, as market demand increases and consumer expectations for delivery rise, the issue of insufficient delivery capacity in e-commerce has become more prominent. This study constructed a delivery path optimisation model that integrates fuzzy time windows. Initially, density-based spatial clustering of applications with noise (DBSCAN) and particle swarm optimisation-based K-means clustering are employed to group customer points by density and identify their centroids, and an improved ant colony optimisation (ACO) algorithm was used to dynamically solve the vehicle routing problem with time window constraints. Experimental results demonstrate that the mean square error of the proposed hybrid clustering algorithm converges to 0.0426 within 139 s, balancing analysis efficiency and clustering accuracy. This model effectively balances computational overhead and optimisation quality, while reducing merchant operating costs and improving customer satisfaction. Keywords: clustering analysis; DBSCAN-IPSO-K-means; ant colony optimisation; ACO; time windows; rolling time domain; delivery route. DOI: 10.1504/IJICT.2026.10079513
Abstract: With the digital transformation of smart education, objective and efficient classroom teaching quality evaluation has become a critical challenge for optimising teaching effectiveness. Traditional manual assessment methods suffer from strong subjectivity, low efficiency, and delayed feedback, while existing AI-driven approaches face bottlenecks of single-modal limitation, insufficient multimodal semantic fusion, and poor scene generalisation. To address these issues, this paper proposes a multimodal transformer for teaching evaluation framework, which deeply integrates speech, text, and behavioural modalities to realise interpretable teaching quality assessment. This framework designs a tri-modal feature extraction module, a cross-modal transformer fusion encoder with adaptive modal weighting. Experimental results show that the model achieves a macro-average F1 score of 0.88 across three core evaluation dimensions, significantly outperforming mainstream baselines. This work provides a robust technical solution for intelligent teaching evaluation and a benchmark for future research. Keywords: teaching quality evaluation; multimodal transformer; cross-modal fusion; speech-text-behaviour. DOI: 10.1504/IJICT.2026.10079514
Abstract: Ethnic dance movement recognition is critical for cultural heritage preservation and intelligent dance teaching, yet traditional methods struggle to capture spatial correlations of human joints and dynamic motion features of dance movements. This paper proposes a recognition method fusing graph convolutional network and motion attention mechanism to address these problems. The graph convolutional network models human skeletal joints as graph structures to extract spatial topological features, while the motion attention mechanism adaptively weights dynamic motion information of different joint sequences to enhance key movement feature representation. Evaluations on public and self-built ethnic dance datasets show the method outperforms traditional convolutional and single graph network models by 8.2% and 5.7% in recognition accuracy, respectively. It effectively captures the unique spatial and dynamic characteristics of ethnic dance movements, providing a reliable technical solution for ethnic dance digitalisation and intelligent analysis. Keywords: ethnic dance movement recognition; graph convolutional network; GCN; motion attention mechanism. DOI: 10.1504/IJICT.2026.10079515
Abstract: This paper addresses the challenge of difficult early warning of muscle injuries during exercise training, proposing a new method that integrates the time-frequency characteristics of surface electromyography signals with biomechanical simulation analysis. Traditional methods cannot comprehensively reflect the functional state of muscles during dynamic movements. This study uses wavelet transformation to deeply explore the latent fatigue and injury precursors in the signals, and simultaneously simulates the mechanical responses of the musculoskeletal system, thereby achieving real-time quantitative assessment of injury risk. Experimental results on public datasets show that compared with traditional time-domain analysis methods, this method increases the accuracy of injury identification from 82.4% to 91.7%, and the area under the curve for risk prediction also improves from 0.86 to 0.93. This method provides a more accurate and reliable decision support tool for monitoring training loads and preventing injuries in athletes. Keywords: surface electromyographic signal; time-frequency analysis; injury risk assessment; biomechanical simulation. DOI: 10.1504/IJICT.2026.10079516
Abstract: As fund management in colleges and universities faces the challenges of low efficiency and intensified operational risks, this study aims to build a new intelligent decision-making model to realise process automation and decision optimisation by integrating robotic process automation (RPA) and artificial intelligence technology. This paper adopts a hierarchical architecture to integrate data layer, process layer, decision layer and application layer, and uses OCR, LSTM and random forest technologies to automatically process invoice identification, fund forecast and risk monitoring. Although the model has achieved remarkable results in improving the efficiency of capital use and reducing risks, it still has limitations such as strong data dependence, and it is necessary to optimise the lightweight and cross-scenario adaptability of the algorithm in the future. Generally speaking, this study provides empirical support for financial intelligence in colleges and universities, and promotes the innovative practice of integrating automation and decision-making. Keywords: colleges and universities; money management; RPA robots; process optimisation; intelligent decision-making. DOI: 10.1504/IJICT.2026.10079517
Abstract: The lack of personalised adaptation in higher vocational curriculum systems is increasingly prominent. To address this, we propose a knowledge graph-based curriculum system construction and path optimisation model (KG-VCSPOM). The model integrates two modules: a knowledge graph construction module using BERT and GGAT to extract course entities and relationships, and a path optimisation module combining GCN embedding and reinforcement learning to compute optimal learning sequences. Tests show the model achieves an F1-score of 84.6% (baseline CF: 71.2%), a noise robustness attenuation rate of 6.2%, and a user satisfaction score of 4.5. Performance remains stable as data scales to 100,000 entries. These results confirm the model's advantages in improving recommendation accuracy, adaptability, and practicality, providing an innovative solution for intelligent higher vocational education. Keywords: knowledge graph; higher vocational education; curriculum system; path optimisation. DOI: 10.1504/IJICT.2026.10079518
Abstract: To reduce word error rate of speech recognition in English interactive system, a key technology of speech recognition based on improved BiLSTM-HMM model is proposed. Firstly, BiLSTM-HMM speech recognition model is constructed by combining BiLSTM network and HMM model. Then, on the basis of BiLSTM-HMM model, attention mechanism is introduced, and model structure and data alignment are improved. Meanwhile, improved BiLSTM-HMM model is proposed, and it is applied to the English interactive system built on voice communication technology. Finally, verification is carried out on LibriSpeech dataset. The results shows that recognition time of improved BiLSTM-HMM model on LibriSpeech dataset is 3,541 s, and its error rate is 0.35; in English interactive system, error rate of English speech recognition for users is 0.34. Therefore, the key technology of speech recognition based on improved BiLSTM-HMM model can be used for speech recognition in English interactive system, and can effectively reduce the error rate of speech recognition in English interactive system, which has certain practical application value. Keywords: English interactive system; speech recognition; BiLSTM network; HMM model; speech communication technology. DOI: 10.1504/IJICT.2026.10079519
Abstract: High-dynamic power optical networks are characterised by frequent topology changes and resource constraints, which challenge traditional routing mechanisms in meeting quality-of-service requirements. This paper proposes a multi-path deep Q-network (DQN) routing mechanism tailored for such environments. By integrating multi-path transmission with deep reinforcement learning, the mechanism optimises delay, packet loss rate, and load balancing through a multi-objective reward function. Built upon a layered graph model of the optical network, the approach incorporates a multi-path selection strategy and employs DQN for intelligent routing decisions under dynamic conditions. Simulation results demonstrate that the proposed mechanism effectively reduces end-to-end delay and packet loss under high traffic loads, while improving resource utilisation and network robustness, making it suitable for efficient data transmission in high-dynamic power optical networks. Keywords: high-dynamic optical networks; multi-path routing; deep Q-network; DQN; deep reinforcement learning; load balancing. DOI: 10.1504/IJICT.2026.10079520
Abstract: Aiming at the problems of inconsistent terminology, strong label coupling, insufficient generalisation, and difficult real-time deployment in power communication customer service texts, this paper proposes a multi-label classification method based on binary relevance (BR) and gradient boosting decision tree (GBDT) with adaptive label correlation weighting. We design an automated training set construction scheme using data cleaning, Jieba segmentation, TF-IDF, and Word2vec to build a standardised label system. The BR-GBDT model decomposes multi-label tasks into binary classification tasks and uses mutual information-based label correlation weights to overcome the label independence defect of traditional BR. Experiments on 12,800 real work orders show that the method achieves Hamming Loss 0.072, Subset Accuracy 82.3%, Macro-F1 89.5%, and Micro-F1 92.1%. The average single-sample prediction time is 18.3 ms under real system load, meeting the 30 ms real-time requirement. The model exhibits strong cross-domain adaptability, noise robustness, and low hardware dependence, and is interpretable via feature importance and per-label performance analysis. Comparisons with CNN-LSTM, BERT, and classic multi-label methods verify its superiority. This method supports efficient, real-time, and deployable multi-label classification for power communication customer service systems. Keywords: power communication; multi-label text classification; binary relevance; BR; gradient boosting decision tree; GBDT; label correlation; model interpretability. DOI: 10.1504/IJICT.2026.10079521
Abstract: A new media public opinion classification model is proposed to address limitations of traditional methods, including high-dimensional sparseness, poor semantic understanding, and category imbalance. The model integrates a three-branch decision feature selection algorithm (combining mutual information and improved term-category information) with a multi-channel random undersampling mechanism. A hybrid architecture combining Transformer, LSTM, and CNN captures global context, sequence dependence, and local features. Experiments on Weibo and Douyin datasets achieve accuracies of 89.47% and 87.93%, with macro F1 values of 83.26% and 81.17%, respectively. The model improves negative opinion recognition F1 by 6.97% over a pre-trained language model, with inference time of 28.5ms (120 samples) and stable 15.3% CPU usage. This work provides an accurate, efficient, and robust engineering solution for new media public opinion monitoring. Keywords: new media public opinion classification; three-way decision feature selection; multi-channel random undersampling; imbalanced data classification. DOI: 10.1504/IJICT.2026.10079522
Abstract: In power communication scenarios, incremental pre-training of large models faces challenges such as limited bandwidth and limited edge device resources , and existing gradient compression methods are difficult to adapt to dynamic network environments. To this end, this paper proposes a power-aware adaptive gradient sparsity compression (PAGSC) algorithm, which achieves efficient training by fusing dynamic sparsity, hierarchical quantisation and channel adaptive scheduling mechanisms. The test results show that on data sets such as CIFAR-100 and UCI power, PAGSC reduces the communication volume to 12% of the baseline, with only 2% accuracy loss, reduces training time by 19%, increases bandwidth utilisation to 85.6%, and significantly optimises resource use. To sum up, the algorithm effectively balances accuracy and efficiency. However, further research on hyperparameter adaptive optimisation and noisy channel robustness is needed. The innovation of this paper lies in providing a deployable lightweight training solution for the power internet of things. Keywords: power communication; large model incremental pre-training; gradient sparse compression. DOI: 10.1504/IJICT.2026.10079523
Abstract: Corporate loan risk prediction is challenged by high-dimensional features, nonlinear relationships and severe class imbalance. To improve prediction accuracy and robustness, an improved LightGBM-based corporate loan risk prediction model is proposed. The framework integrates D-B-SMOTE for minority sample balancing, a feature importance fusion selection mechanism for adaptive feature screening, and a hybrid loss function combining focal loss and Huber loss to enhance sensitivity to default samples and reduce noise interference. In addition, an SSA-GWO hybrid optimisation strategy is introduced for hyperparameter tuning, and a distributed parallel training architecture is constructed to improve scalability. Experimental results show that the proposed model achieves an accuracy of 0.893, recall of 0.822, F1-score of 0.837 and AUC of 0.886, demonstrating strong robustness and generalisation ability in enterprise loan risk prediction. Keywords: corporate loans; feature fusion screening; LightGBM; risk prediction; sparrow search algorithm; SSA. DOI: 10.1504/IJICT.2026.10079524
Abstract: To improve the stability of distributed collaborative media data transmission, this paper proposes a two-stage distributed collaborative media transmission optimisation method based on improved K-means algorithm and improved bee evolutionary genetic algorithm (BEGA). By dividing distributed collaborative media transmission optimisation into clustering and routing stages, and respectively using K-means algorithm improved by introducing a differential evolution algorithm, as well as BEGA algorithm improved in parameters, crossover method, and mutation method for clustering and optimisation, the optimisation of distributed collaborative media transmission is achieved. The simulation results show that when number of deployment node is 100, 300, and 500, number of surviving nodes drops to zero approximately after 700 rounds, the packet transmission rate is about 70% after 1,000 rounds, the network remaining energy drops to zero approximately after 1,000 rounds, and network energy consumption balance remains at a high value before 800 rounds. Compared with the contrast method, it has more efficient and stable data transmission performance. From this, it can be concluded that the proposed method can improve the stability of distributed collaborative media data transmission, providing a reference for achieving more efficient and stable data transmission in distributed collaborative media. Keywords: distributed collaborative media; transmission optimisation; K-means algorithm; BEGA algorithm. DOI: 10.1504/IJICT.2026.10079525
Abstract: Teaching Chinese tea culture to international students remains difficult in a conventional classroom because key elements of tea practice, including utensil handling, brewing sequence, and the cultural atmosphere of the tea room, are hard to reproduce authentically. This study presents the design and classroom implementation of an immersive virtual simulation learning system based on a browser/server architecture to address this gap. The system integrates three learning scenes, namely the tea culture corridor, tea garden roaming and tea room experience, and organises procedural tasks through a finite-state-machine interaction model. Multimodal guidance, including voice explanation, subtitles, icons, and real-time prompts, is provided to support learners from diverse linguistic and cultural backgrounds. The system was applied in three rounds of teaching involving 300 international students. Classroom records showed stable completion of tea-making tasks, effective retention of core cultural knowledge, and comprehensive learning logs for instructional review. Compared with the baseline classroom condition, the virtual simulation classes demonstrated higher engagement, task completion, and satisfaction at the descriptive level. These findings suggest that browser-based virtual simulation can serve as a practical instructional approach for cross-cultural tea culture education in universities. Keywords: virtual simulation; cross-cultural education; learning system; human-computer interaction; learning behaviour analysis. DOI: 10.1504/IJICT.2026.10079531
Abstract: To address the problems of low prediction accuracy and poor real-time performance in traditional cable tunnel construction behaviour prediction methods, a prediction method based on Qwen3-VL temporal enhancement is proposed. First, multi-source data related to cable tunnel construction (including construction equipment operation parameters, environmental monitoring data and process execution records) are collected and preprocessed to eliminate noise and data redundancy. Then, the temporal enhancement module of Qwen3-VL is optimised to enhance the models ability to capture time-series features of construction behaviour, effectively mining the temporal correlation and evolution rules among multi-dimensional construction data. Finally, the enhanced feature set is input into the prediction network to realise accurate prediction of key construction behaviours such as non-standard operation, process deviation and equipment failure risk. Experimental results show that compared with traditional machine learning methods and general deep learning models, the proposed method improves the prediction accuracy by 8.3%-12.7% and shortens the prediction response time by 15%-22%, which can provide reliable technical support for safe and efficient construction of cable tunnels. Keywords: cable tunnel; construction behaviour; prediction optimisation; Qwen3-VL; temporal enhancement. DOI: 10.1504/IJICT.2026.10079587
Abstract: As global population aging intensifies, precise prediction of elderly care needs is of vital importance for the optimal allocation of medical resources and personalised intervention. This paper proposes a hybrid framework that integrates spatial-temporal transformer and deep reinforcement learning to predict multiple care needs using longitudinal health and social data. This framework models the long-term dependencies of health trajectories through spatial-temporal Transformer and cross-modal interactions among clinical, functional, and social factors. It also incorporates a reinforcement learning component to optimise the sequential decision-making process of care need evolution. Experiments on two datasets show that the proposed method achieves significant improvements over nine baseline models and four evaluation metrics. Specifically, the AUC increases by 3.2% to 7.8%, and the F1-score increases by 4.1% to 9.3%. This research provides a new methodological tool and empirical basis for elderly care planning. Keywords: care needs for the elderly; spatiotemporal transformer; deep reinforcement learning. DOI: 10.1504/IJICT.2026.10079678
Abstract: The dynamic evolution and nonlinear characteristics of stock market risks pose significant challenges to traditional forecasting methods. To overcome the limitations of point forecasting, this paper introduces risk simulator a deep learning framework integrating temporal encoding, state transition mechanisms, and multi-task learning to achieve temporal simulation and multi-step prediction of risk states. By explicitly modelling the evolutionary paths of risk levels, this model enhances interpretability and prediction accuracy. Experiments on four public datasets, including the standard and poors 500 and China securities index 300, demonstrate that risk simulator achieves an area under the curve of 0.831 - a 3.2% improvement over the state-of-the-art benchmark risk-aware temporal network - and a prediction accuracy of 0.779, representing a 2.3% relative gain. This research provides a new simulation-based tool for quantitative risk management, validating the effectiveness of multi-task learning and state evolution mechanisms in financial time series modelling. Keywords: stock market risk; deep learning; time-series simulation. DOI: 10.1504/IJICT.2026.10079679
Abstract: Facing global cross-language interaction scenarios, the demand for high-quality English machine translation is increasingly urgent. However, current research suffers from issues such as lack of structured knowledge and long-text semantic fragmentation. To tackle these issues, a method is proposed in this paper that leverages both bidirectional encoder representations from transformers (BERT) and a knowledge graph to improve English text translation quality. An improved TransR model is employed to extract entity embeddings from the knowledge graph, obtaining knowledge-enhanced representations. Multi-layer BERT outputs are fused and a masked knowledge matrix is constructed to suppress semantic noise. An adaptive fusion module is designed to dynamically integrate BERT knowledge with encoder context, jointly optimising translation and knowledge losses. Experimental results show that BTKG achieves BLEU scores of 34.6 and 33.2 on EnglishGerman and EnglishChinese translation tasks, with translation accuracies of 92.7% and 90.1%, respectively, significantly outperforming existing baseline methods. Keywords: English text translation; bidirectional encoder representations from transformers; BERT; knowledge graph; knowledge embedding; adaptive fusion. DOI: 10.1504/IJICT.2026.10079680
Abstract: Text provenance aims to determine the authorship of an anonymous text, which is crucial in the field of digital forensics and academic integrity. Existing models are difficult to deal with hierarchical style features of long documents, and cannot run in real-time on edge devices. This paper proposes edge-empowered hierarchical attention network, which integrates the cognitive load theory, extracts writing habits through the word-sentence-paragraph-full text layer-by-layer attention mechanism, and designs constraints to distinguish scene differences without confusing author identities. Finally, the model is compressed to the Edge end by pruning and distillation. On the PAN 2020 dataset, the traceability accuracy of the proposed framework reaches 0.923, which is 4.2% higher than the current best method. The edge-side inference takes only 12.3 Ms. This paper quantifies cognitive load theory as a computable constraint for the first time, and provides a high-precision solution for real-time text provenance at the edge. Keywords: text provenance; hierarchical attention; cognitive load theory; edge computing; authorship attribution. DOI: 10.1504/IJICT.2026.10079681
Abstract: Cross-era wind instrument performances are often accompanied by significant rhythmic distortions and the insertion of ornamental notes; traditional dynamic time warping struggles to handle such structural mismatches. This paper integrates transformers with weighted dynamic time warping: the transformer encoder captures long-range dependencies in the musical score to generate enhanced rhythmic features; weighted dynamic time warping dynamically adjusts the cost of local matching via a learnable weight matrix to achieve flexible alignment. on the musical instrument digital interface and audio edited for synchronous tracks and organisation dataset and a custom-built wind instrument subset, our method achieves a precision of 86.2% (a 7.7% improvement over dynamic time warping), a recall of 83.1% (a 6.2% improvement), and an normalised discounted cumulative gai @5 of 0.723 (a 12.8% improvement). The study demonstrates that the neural-symbolic fusion strategy can effectively decompose cross-epochal rhythmic variations in wind music, providing new insights for music information retrieval. Keywords: transformer architecture; dynamic time warping; DTW; wind instrument rhythm decomposition; cross-generational performance. DOI: 10.1504/IJICT.2026.10079682
Abstract: Students often fail to align course selection with long-term career goals due to overwhelming resources and lack of career-oriented guidance. Existing learning path methods emphasise short-term performance but neglect vocational skill accumulation, resulting in suboptimal career preparation. This paper proposes a knowledge graph-enhanced reinforcement learning framework generating personalised course sequences for career advancement. Grounded in social cognitive career theory, it integrates self-efficacy, outcome expectations, and personal goals into state and reward mechanisms. A knowledge graph integrating courses, skills, and occupations from occupational information network is embedded via a graph convolutional network to provide skill-aware student representations. A deep Q-network optimises sequences by maximising a dual-objective reward balancing grades and career alignment. Experiments show this framework achieves a 68.5% career achievement rate, outperforming the strongest baseline by 14.2 percentage points, while reducing average courses to reach a career goal to 5.6. Keywords: knowledge graph; reinforcement learning; personalised learning path; career development; social cognitive career theory. DOI: 10.1504/IJICT.2026.10079683
Abstract: In todays rapidly evolving marketing landscape, promotional campaigns have become a vital strategy for businesses to expand their market reach. However, the explosive surge in user traffic during promotions often poses significant challenges to corporate system resources, leading to degraded system performance. To address this, this paper first outlines a rational planning approach for all resources, optimising their allocation. Subsequently, based on user requests, the proxy server selects pre-configured resources from the resource pool and allocates them to users. Throughout the resource scheduling process, dynamic allocation and capacity planning for promotional resources are achieved using linear programming. This approach targets the total utility value of completed user request tasks while constraining expected task duration, costs, and parallel acceleration ratios. The experimental findings indicate that the proposed system achieves 87.8% usage of computational resources, 85.2% of storage resources, and 93.1% of network bandwidth resources. Keywords: market promotion; resource allocation; capacity planning; utility computing linear programming. DOI: 10.1504/IJICT.2026.10079684
Abstract: Complying with data privacy rules across national borders has become a central obstacle for the effective conduct of cross-border business. Traditional methods suffer from low accuracy and are vulnerable to attacks initiated by malicious actors. To address this, this paper first proposes a cross-border privacy data compliance verification model based on federated learning and generative adversarial networks. By employing a generator and a discriminator, this model learns the data distributions of local nodes distributed across multiple countries. Generator parameters are then uploaded to the global coordinator in place of model parameters. Next, a dynamically authenticatable federated learning mechanism is constructed to effectively defend against backdoor attacks. Lastly, homomorphic encryption is integrated to encrypt the generator parameters of nodes, ensuring the preservation of node privacy. Experimental results demonstrate that under a 60% proportion of malicious nodes, the proposed model achieves a verification accuracy of 98.6%. Keywords: cross-border data privacy; dynamic compliance verification; federated learning; FL; generative adversarial network. DOI: 10.1504/IJICT.2026.10079685
Abstract: Traditional measures of college students values rely on self-report questionnaires, which are susceptible to social desirability effects and introspective bias, making it difficult to capture genuine, unconscious tendencies. To address this challenge, this paper proposes a multimodal affective computing-driven implicit measurement paradigm. It collects students microexpressions, speech prosody, and eye-tracking data in response to value-conflicting stimuli, and constructs a temporal attention network to uncover the mapping patterns between emotional responses and deep-seated value orientations. The experiment utilised both publicly available emotional datasets and a custom-built database. Results showed that the multimodal fusion model achieved an area under the curve of 0.87 in classification tasks a 14% improvement over single-visual-modal approaches and a normalised cumulative gain of 0.91 in ranking prediction, validating the validity advantages of implicit measurement. This study provides a feasible pathway to overcome the limitations of self-reporting and achieve objective assessment of values. Keywords: multimodal affective computing; latent measures; college students’ values; physiological and behavioural responses. DOI: 10.1504/IJICT.2026.10079686
Abstract: Designing the optimal parameters for an integrated photovoltaic-storage-charging (PVSC) station is a complex task. Traditional methods rely on simulating all possible parameter combinations, which is time-consuming and computationally intensive. To accelerate the optimisation process and reduce computational costs, this paper proposes a novel approach combining generative adversarial networks (GANs) and reinforcement learning. First, the performance of a subset of parameter combinations is evaluated through simulation of charging demand, and these data are used to train the GAN to generate additional synthetic samples, thereby expanding the search space. Then, an artificial neural network (ANN) surrogate model is constructed to predict the performance metrics for any given parameter combination. Finally, the surrogate model is used within a reinforcement learning framework to efficiently search for the optimal parameters. Experimental results show that this method effectively accelerates the design optimisation process, providing an efficient framework for quickly determining the optimal parameters of photovoltaic charging stations. Keywords: integrated photovoltaic-storage-charging; charging station; reinforcement learning; neural networks; design parameter optimisation. DOI: 10.1504/IJICT.2026.10079718
Abstract: Rhythm analysis of English texts is a core task in fields such as speech synthesis and affective computing. However, traditional methods often focus on single-modal data, making it difficult to capture deep semantic relationships across modalities. To address this, this paper first employs deep learning algorithms to extract multimodal English features containing significant semantic information. Subsequently, a hierarchical semantic enhancement module is designed to fuse relevant semantic information from visual and audio modalities at different text levels. Finally, we propose a multimodal feature fusion module based on cross-modal attention and gating mechanisms to suppress relatively unimportant features, thereby reducing redundant information and improving the accuracy of English text rhythm classification. Experimental results demonstrate that the proposed method achieves a weighted average F1 score of 93.17%, significantly outperforming comparison methods and enabling precise English text rhythm analysis. Keywords: English text; rhythm classification; semantic enhancement; multimodal fusion; deep learning. DOI: 10.1504/IJICT.2026.10079719
Abstract: In this study, an innovative model fusing reinforcement learning and U-Net is proposed to address the demand for automated analysis of childrens painting quality. The model extracts multi-level visual features of paintings using an enhanced U-Net and introduces a reinforcement learning agent to simulate the sequential decision-making process of expert evaluation, thereby enabling comprehensive analysis of multidimensional quality indexes such as composition, line, colour, and integrity. The experiment is based on a self-built childrens painting data set containing 15,832 labelled samples. The model achieves an evaluation accuracy of 92.6% on the test set, and its Spearmans rank correlation coefficient is 0.891, which is significantly better than those of the traditional CNN method and the basic U-Net model. The ablation experiment further shows that introducing a reinforcement learning reward mechanism reduces the models average absolute error in the composition rationality analysis task by 0.15, effectively improving the stability and interpretability of the evaluation results. This study verifies the models effectiveness for analysing childrens painting quality and provides a new technical path for intelligent evaluation of aesthetic education. Keywords: reinforcement learning; U-Net algorithm; children’s drawing analysis; image quality evaluation; intelligent education. DOI: 10.1504/IJICT.2026.10079720
Abstract: With the boom in educational informatisation and online learning platforms, the mismatch between educational resource supply and learners personalised needs has become acute, hampering resource allocation efficiency and learning outcomes. This study proposes a deep learning-based educational resource recommendation system for demand-supply matching. First, we model educational resource features and learning needs, then design a deep learning matching model integrating multi-source features, and formulate personalised recommendation strategies using matching probabilities. Depending on the experimental verification, it can be found that the proposed method achieves a relatively stable performance in the task of supply and demand matching, and its F1-value reaches 0.829, which improves its performance by about 17% compared with the traditional collaborative filtering method. Ablation experiments confirm that incorporating learner and resource features significantly boosts model performance, and the method exhibits superior hit and resource completion rates, verifying its effectiveness in enhancing resource allocation efficiency. Keywords: deep learning; educational resources; supply and demand matching; personalised recommendation; recommendation system. DOI: 10.1504/IJICT.2026.10079721
Abstract: This paper proposes a personalised recommendation algorithm based on big data analysis. The algorithm integrates multi-dimensional student information and optimises collaborative filtering methods to achieve efficient matching of translation learning resources. Firstly, a data collection system is established to gather student behaviour, achievements, resources, and learning environment data. Secondly, a dynamic analysis model is developed to integrate learner characteristics and cognitive level information. Finally, a translation opportunity comparison function is added to enhance resource accuracy and address data scarcity issues. The mean absolute error achieved 0.152, with the cumulative gain in normalised loss for the top 10 recommendations reaching 0.834. Practical application yielded significant outcomes: following 60 students use of the recommendation system, average translation scores rose from 65.3 to 82.7 points, representing a 26.6% increase. Training engagement markedly improved, with session duration extending by 71.2% and interaction rates surging by 151.6%. Keywords: big data analysis; translation teaching; personalised recommendation; collaborative filtering; learner portrait; education big data. DOI: 10.1504/IJICT.2026.10079722
Abstract: This study integrates a backpropagation neural network optimised by particle swarm optimisation with a two-stage least squares causal inference framework to analyse household data from the China Household Finance Survey. The approach mitigates endogeneity from reverse causality. Results indicate that the digital economy expands household financial consumption scale and optimises consumption structure. Mechanism tests show that household property income and household head financial literacy mediate this effect, with a sequential mediation pathway from financial literacy to property income. Heterogeneity analysis reveals stronger effects in regions with advanced financial development and among risk-preferring households. These findings provide empirical evidence for the role of the digital economy in consumption upgrading. Keywords: digital economy; household financial consumption; PSO-BP models; causal inference; heterogeneity detection; CHFS data. DOI: 10.1504/IJICT.2026.10079723
Abstract: Static scheduling inadequately addresses dynamic tourist preferences and spatiotemporal resource heterogeneity in cultural tourism. This study proposes a hybrid framework integrating AlphaEvolve with graph convolutional networks (GCNs). The model extracts topological features via GCNs and constructs dynamic strategy spaces through AlphaEvolve, learning resource-demand matching from real-time behavioural and environmental data. Results demonstrate 82.7% overall recommendation accuracy (89.6% historical, 82.4% real-time, 76.1% dynamic event scenarios). Response time decreases from 12.5 to 4.8 minutes, with 91.7% peak-hour resource utilisation and 3.1% wait timeout rate. User satisfaction improves substantially: content relevance (71.2 89.5), timeliness (63.8 91.2), short-term conversion (12.3% 28.7%), long-term conversion (8.9% 22.4%), and daily usage reaches 32.1 minutes. Strategy evaluation error converges below 0.1, enhancing scheduling stability by 33%. Keywords: cultural and tourism resource scheduling; AlphaEvolve algorithm; graph convolutional network; GCN; dynamic recommendation; multi-objective optimisation. DOI: 10.1504/IJICT.2026.10079786
Abstract: Against the backdrop of educational digital transformation, traditional short-term technical training for teachers AI application skills suffers from poor teaching scenario transferability and diminishing long-term effectiveness due to unclear internal skill formation mechanisms. This study constructs an intervention system integrating causal inference (to identify key influencing factors) and reinforcement learning (to deliver dynamic personalised interventions). A 4-week experiment shows the intervention group achieved an 11.5-point comprehensive improvement (vs. 4.2 points in the control group), with significant gains across all skill dimensions and higher subjective satisfaction. This dual-driven model overcomes static training limitations, providing methodological support for sustainable teacher AI skill development and educational digital transformation. Keywords: causal inference; reinforcement learning; teachers’ AI application abilities; dynamic intervention; long-term improvement. DOI: 10.1504/IJICT.2026.10079787
Abstract: This study develops a systematic framework integrating outcome-based education (OBE) and cognitive psychology to align digital media technology teaching with industry needs. Using a mixed-methods approach, we established a demand-driven-cognitive alignment-continuous improvement (D-C-I) mechanism featuring a five-domain competency matrix, Scrum agile methods with cognitive load optimisation, and a multimodal data fusion platform for closed-loop feedback. Empirical results showed significant improvements: 3ds Max proficiency increased by 34.4%, project completion rates rose from 71.4% to 94.4%, VR course achievement improved from 68% to 85% (z = 3.87, p < 0.001), and professional alignment rates increased from 61% to 93% (2(1) = 24.67, p < 0.001). The goal-process-evaluation model realises a paradigm shift from static standards to dynamic algorithms, providing a replicable engineering paradigm for higher vocational education reform. Keywords: OBE philosophy; cognitive psychology; D-C-I mechanism; digital media technology; teaching standards; dynamic reform. DOI: 10.1504/IJICT.2026.10079788
Abstract: In the field of basketball biomechanical analysis, traditional methods often rely on single sensors or two-dimensional videos, making it difficult to comprehensively capture the complex spatiotemporal characteristics of shooting movements. This study proposes a multi-sensor fusion analysis method based on spatiotemporal graph convolutional networks (ST-GCN), aiming to achieve accurate and automated analysis of shooting movements. Specifically, a novel multi-stream fusion ST-GCN model architecture is first constructed, which can simultaneously process heterogeneous time series data from inertial measurement unit (IMU) and motion capture systems. Secondly, a cross-modal data fusion strategy based on attention mechanism is proposed, effectively enhancing the representation ability of key features such as joint angles and motion trajectories. Finally, end-to-end motion recognition and synchronous analysis of biomechanical parameters (such as joint torques and angular velocities) are achieved. Experiments collected multimodal data from over 2,000 shooting movements of 50 athletes. The results show that the method achieves an action classification accuracy of 96.7% on the complete dataset, and an accuracy of 92.5% in the IMU+sEMG sensor combination verification. When predicting key biomechanical parameters, the correlation coefficient with professional force platform data is as high as 0.93, significantly outperforming traditional analysis methods based on rules or single sensors. Keywords: spatiotemporal graph convolutional network; ST-GCN; multi-sensor fusion; shooting action analysis; biomechanics; attention mechanism.
Abstract: In the scenario of long-term English writing assessment, learners vocabulary output characteristics are susceptible to the continuous change of time span and context, which leads to the distortion of vocabulary richness indicators under static assessment. We design a dynamic context-aware outlier detection mechanism based on graph autoencoder, which aims to eliminate the interference of static word frequency preference and restore the learners real dynamic vocabulary development trajectory. We use the sliding window-based depth map clustering technique because it can associate isolated words into a structured semantic network, so as to accurately identify those outliers that appear abrupt or completely incoherent depending on the context. Discarding a general qualitative analysis, our extensive experiments lead to specific numerical verification. Compared with the latest dynamic isolation forest benchmark model, the proposed mechanism improves the comprehensive evaluation index score to 92.4%, which proves that it has high stability and processing efficiency. Keywords: dynamic outlier detection; vocabulary richness; graph autoencoder; GAE; English writing assessment; spatiotemporal sequence modelling.
Abstract: This study develops a performance-based green building assessment system (PBGBAS) centred on energy performance and supported by a configurable benchmark database. The framework integrates input, evaluation, adjustment, output and interpretation modules, enabling calibration by region, building type and lifecycle stage. Indicators are screened from major rating tools and reorganised into a compact hierarchy of performance categories, items and indicators. Measured or simulated values are mapped to a five-point score through transparent linear or nonlinear scoring functions, and weights are obtained from structured stakeholder elicitation. The adjustment module updates indicators, benchmarks and weights to improve system extensibility. A residential eco-city case demonstrates the method. Energy performance achieves the highest score (9.7), whereas future value is lowest (8.4); environmental protection, health and user value score 9.4, 9.3 and 8.9, respectively. Results indicate that PBGBAS supports more operational, adaptable and performance-oriented green building evaluation. Keywords: performance-based assessment; energy performance; benchmark database; modular rating tool; regional calibration; PBGBAS.
Abstract: In university ideological and political education, students ideological states are dynamic and hard to measure directly. The traditional uniform teaching model cannot achieve individualised instruction, and existing big data research lacks operational precise mechanisms and empirical verification. To address this gap, this paper proposes a hybrid-driven framework integrating fuzzy c-means clustering, random forest, and long short-term memory network. By constructing dynamic student profiles and context-aware recommendation strategies, it enables personalised resource distribution. Experiments show the model increases prediction accuracy from the random forest baseline of 86.2% to 89.5%, and area under the curve from 0.891 to 0.927. The introduced group response consistency index and precise distribution uniformity reach 0.842 and 0.763, respectively, outperforming the uniform distribution strategy without personalised recommendations. This framework fills the empirical gap from theoretical conception to quantifiable deployment. Keywords: big data; ideological and political education; precision education; hybrid learning model.
Abstract: The automatic correction of English pronunciation has long been hindered by two major problems: extremely scarce non-native annotated data, and the difficulty for traditional models to effectively capture the pronunciation similarities between phonemes. To address these issues, this paper proposes a pronunciation correction framework named pronunciation correction with graph and transfer, which structures the phonemes into a relational graph to utilise its dependency information, and leverages cross-language transfer learning to reduce reliance on target annotated data. On public datasets, this method achieves area under the curve scores of 0.862 and 0.879 respectively, which are 7.3% and 6.8% higher than the baseline model; in precision metrics, it achieves improvements of 8.1% and 7.4%. The results show that pronunciation correction with graph and transfer significantly enhances the accuracy of pronunciation error detection, providing a feasible path for computer-assisted pronunciation training in data-scarce scenarios. Keywords: pronunciation correction; graph neural network; transfer learning; phoneme relationship graph.
Abstract: Short videos have become a dominant medium for ideological and political education among young people, yet many convey hidden value biases through mismatched visual, acoustic and textual information, silently compromising educational effectiveness. To automatically detect this value conduction deviation, this paper proposes the multimodal value deviation network integrating slow fast, wave-to-vector 2.0 and bidirectional encoder representations from transformers with a modality-gated attention mechanism. This paper introduces a modality contribution regularisation loss and a contrastive alignment loss to enhance detection accuracy and interpretability. After re-annotating the visual-audio-text values dataset under the Chinese ideological education context, our method achieves an F1 score of 0.862 and an area under the curve of 0.904 on the test set, outperforming the best baseline by 11.2% in F1 score with statistically significant improvement. The framework also identifies the dominant modality causing the deviation, demonstrating practical value for automated content moderation. Keywords: value conduction deviation; multimodal perception; short video; ideological and political education; interpretable assessment.
Abstract: To address the issues of high computational cost in the routing simulation of the field-programmable gate array physical design process and the neglect of the global topological structure of the circuit, a method for predicting the routability of integrated circuits based on deep learning is proposed. Circuits are abstracted into complex networks, from which topological features are extracted. These are fused with traditional circuit features and mapped onto heatmaps, achieving unified representation of geometric and topological information. Subsequently, the routing simulation task is transformed into an image generation problem. By introducing an efficient channel attention mechanism within a conditional generative adversarial network framework, the method highlights crucial feature channels to enhance generated image quality. Experimental results demonstrate that the proposed approach achieves 98.26% structural similarity and a peak signal-to-noise ratio of 41.75 dB, both significantly outperforming existing methods. This method provides a new paradigm for rapid assessment of routability prediction during the layout stage. Keywords: field-programmable gate array; FPGA; routing process simulation; attention mechanism; complex network; generative adversarial network; GAN.
Abstract: To address the lack of systematic classification and trend analysis for creative packaging pattern design styles, this paper proposes a novel multimodal deep clustering framework. The method integrates convolutional neural network texture features, vision transformer global structure features, and simple linear iterative clustering balanced iterative reducing and clustering using hierarchies colour features, fusing them adaptively with learnable weights for dynamic feature fusion and end-to-end jointly optimised with deep embedding clustering to achieve unsupervised style classification. On the self-built packaging dataset containing 9,000 images, the method consistently achieves an accuracy of 0.768 and an F1-score of 0.755, thus significantly outperforming existing approaches (p < 0.05). The results reveal design evolution trends from 2015 to 2024, providing valuable data-driven support for design decision-making, trend prediction, and practical design practice. Keywords: clustering algorithm; packaging design; style classification; trend analysis; multimodal features.
Abstract: This study tackles inefficiencies in virtual drama character action recognition. An improved skeleton extraction method employs sparse matrix operations and a novel activation function to reduce parameters and enhance nonlinearity. The core model uses an enhanced spatial-temporal graph convolutional network with a learnable matrix and multi-head graph attention to resolve topological rigidity. A twin network enables precise action quality assessment. Experiments demonstrate the superior performance. It achieves 97.8% average skeletal feature recognition accuracy across dramatic scenes and 95.85% accuracy in specific scenarios such as ballet jumps, outperforming comparable models. It also maintains robustness in noise, reaching 81.2% accuracy at a 0.25 noise level. For quality assessment, its evaluations align closely with expert judgements, achieving 98.6% average accuracy in rhythm matching tasks. In summary, this technique provides effective technical support for optimising virtual drama character action systems. Keywords: virtual drama; action recognition; ST-GCN model; attention mechanism; skeletal features.
Abstract: Vocal emotional expression is a highly complex, time-varying process. Current generative models often struggle with long-term dependencies and fine-grained dynamic transitions, causing over-smoothed acoustic features. To address this limitation, a temporal-conditioned diffusion model is proposed for the high-fidelity simulation of time-varying vocal emotional features. The model utilises a stochastic differential equation based diffusion process operating on a multidimensional acoustic feature space. During the reverse generative phase, a hybrid LSTM-Attention network is employed to capture both macroscopic and microscopic emotional transitions. Comprehensive evaluations demonstrate that the proposed model significantly outperforms baselines. specifically, our model achieved a lower Mel-cepstral distortion (3.95 dB), reduced F0 RMSE (12.8 Hz), and a higher energy Pearson correlation (0.92). Furthermore, subjective mean opinion scores for acoustic naturalness (4.35) and emotional expressiveness (4.42) surpassed the best baseline. This study provides a robust generative paradigm for advanced affective computing. Keywords: vocal emotional characteristics; time-varying process; deep generative model; diffusion model; simulation modelling; emotional computing.
Abstract: Academic performance prediction is essential for personalised teaching intervention, yet electromechanical courses involve complex precedence dependencies and logical correlations between knowledge points that traditional models fail to capture. To address this, we propose a learning-performance prediction method named knowledge-graph-based mechatronics education learning prediction, which integrates expert knowledge with historical interaction data to construct a knowledge graph encompassing 56 core knowledge points. This method employs graph convolutional networks (GCNs) to integrate topological features for representation learning and introduces a temporal-attention mechanism to model the dynamic evolution of knowledge mastery. Experimental results demonstrate that the proposed method achieves an area under the curve of 0.894, which is 5.2% to 11.7% higher than baseline models, and a normalised discounted cumulative gain at 5 of 0.856, significantly outperforming comparative methods. The findings highlight the effectiveness of incorporating knowledge-graph structures for accurate academic performance prediction in electromechanical education. Keywords: academic performance prediction; knowledge graph; graph convolutional network; GCN; temporal attention.
Abstract: In complex retail systems, existing process models rely on static historical data, failing to capture the nonlinear dynamics of fluctuating consumer emotions throughout the market lifecycle. This paper proposes the emotion-decision synergistic simulation network, a novel simulation framework for lifecycle strategy optimisation. First, a deep extraction module using bidirectional long short-term memory and multi-head attention models the temporal evolution of emotional states from unstructured textual feedback. These continuous state vectors are then integrated into a multi-agent environment, explicitly simulating the dynamic state transitions and behavioural processes of virtual consumers under marketing stimuli. Extensive experiments across three large-scale datasets demonstrate that EDSSN significantly enhances process evaluation, reducing the mean absolute percentage error by up to 4.15% compared to traditional baselines. Ultimately, ours model provides a robust simulation testbed, successfully bridging qualitative human factors and quantitative process modelling for reliable decision support. Keywords: multi-agent reinforcement learning; bidirectional long short-term memory; Bi-LSTM; multi-head attention mechanism; demand forecasting; consumer sentiment.
Abstract: Oil painting emotion recognition faces two challenges: the excessive number of model parameters makes deployment difficult, and the granularity of emotion semantic understanding is coarse. To address this issue, this paper proposes a lightweight representation learning method that integrates knowledge distillation and multi-granularity semantic alignment. The teacher network guides the lightweight student network to learn, and performs feature alignment at two granularities: global image semantics and local emotional attributes, enabling the compact model to accurately capture the subtle emotions in the paintings while maintaining efficiency. Experiments show that, with a 84.7% reduction in parameters, the classification accuracy only drops by 1.2%, and the normalised discounted cumulative gain for emotion intensity ranking reaches 0.893, significantly outperforming the lightweight baseline models of the same scale. This research provides a technical solution that balances efficiency and accuracy for the practical implementation of emotion computing in the field of digital humanities. Keywords: knowledge distillation; KD; multi-granularity semantic alignment; oil painting sentiment analysis; sentiment computing.
Abstract: Financial statement fraud often hides behind locally plausible numbers while breaking cross-statement coherence, which weakens traditional ratio screens and single-source detectors. To address this challenge, a dual-layer knowledge graph and graph-driven fraud detection framework is proposed. In the framework, first, accounting semantics and consistency constraints are organized as a schema layer and linked to firm-year observations as an instance layer. Then, layer-coupled graph learning aligns observations with accounting meaning and extracts compact evidence subgraphs for audit review. Finally, a generative consistency module estimates coherent reporting profiles to expose distributed manipulation through residual inconsistency. Experimental results show that the proposed method achieves an average precision of 0.74 and improves low-false-alarm recall to 0.60, outperforming representative baselines by up to 0.14 in average precision. The framework provides reliable risk ranking and practical interpretability for audit-oriented deployment. Keywords: financial statement fraud; dual-layer knowledge graph; heterogeneous graph learning; evidence subgraph extraction.
Abstract: Supply and guarantee chains among enterprises create complex networks with significant risk transmission effects. Traditional early warning models relying on isolated financial indicators cannot capture this dynamic evolution, causing delayed and inaccurate warnings. This paper proposes temporal knowledge graph dynamic graph neural network, a real-time framework integrating temporal knowledge graphs and dynamic graph neural networks. It constructs a multi-relation temporal knowledge graph containing enterprises, transactions, and guarantees to depict risk transmission paths, and designs a time-aware graph attention mechanism. Experiments on public datasets show the proposed method achieves 0.962 area under the curve, 3.2% higher than graphshield, and 0.781 normalised discounted cumulative gain at position 10, 8.7% higher than dynamic graph neural network with static relationships. Results demonstrate this technology effectively identifies early risk signals in network environments, providing timely and interpretable decision support for business and finance risk control. Keywords: temporal knowledge graph; dynamic graph neural network; business and finance risk; early warning.
Abstract: Traditional assessments of student legal literacy rely on static questionnaires or performance analysis, which often fail to capture structural complexity and predict highrisk behaviours accurately. To address this, we propose an integrated algorithm combining logistic regression (LR) and a latent structural model (LSM) based on the twoparameter logistic IRT model. LSM estimates latent traits across legal knowledge, concepts, and abilities, mitigating nonlinearity and measurement error, while LR incorporates demographic, behavioural, and latent trait data to predict future highrisk behaviour. A joint optimisation mechanism quantifies literacy and enables early warning of risks. Experiments show the fusion model achieves 88.5% accuracy in highrisk behaviour prediction, outperforming standalone LR (78.2%) and structureddataonly models (75.5%). It also yields an F1 score of 0.875 and recall of 86.0%, demonstrating robust performance. The approach offers a refined framework for legal literacy assessment and supports targeted prevention strategies in educational management. Keywords: legal literacy assessment; logistic regression; latent structural model; LSM; latent traits; high-risk prediction.
Abstract: This study addresses the mismatch between proliferating multimodal resources and learners cognitive needs in digital education. Conventional manual annotation and static classification fail to capture cross-modal semantic associations or adapt to evolving cognitive levels. We propose a knowledge forest-transformer model integrating hierarchical knowledge organisation with self-attention mechanisms for multimodal semantic association and cognitive adaptation. Structured multimodal feature modelling improves cross-modal accuracy: text-image from 68.2% to 89.7%, text-audio from 59.8% to 83.6%, and image-video to 90.2%. Cognitive adaptation achieves 85.8% success at threshold 1.8, while cognitive labels enhance F1-scores (analysis layer: 61.2% to 79.5%; application layer: +17.8%). Response times under three seconds yield cognitive load of 4.2 (12.5% exceedance) versus 8.4 (68% exceedance) when exceeding eight seconds. After four weeks, low-ability learners improve mastery from 32.1% to 61.3%, with medium and high-ability groups reaching 82.7% and 94.2%, respectively. Results demonstrate effective multimodal resource association and cognitive adaptive support. Keywords: multimodal teaching resources; semantic association; cognitive level adaptation; knowledge forest; transformer.
Abstract: To address the challenges of structural coherence and stylistic encoding in ethnic music generation, this paper proposes a novel three-dimensional sequence convolutional network (3D-SCN). The model integrates the translational invariance of CNNs, the sequence modelling of bi-LSTMs, and the feature fusion of multi-head attention. Through multi-dimensional feature extraction and spatiotemporal modelling, it captures patterns in traditional Chinese pentatonic music. Experiments show the model outperforms baselines like music transformer, achieving 98.7% accuracy, a 0.941 F1-score, and 95.6 in expert-rated ethnicity. It also demonstrates strong cross-cultural adaptability (91.3% cultural accuracy) and long-sequence generation capability (92.8% structural integrity). Despite computational complexity, this work provides an effective framework for digital music generation and heritage. Future work will focus on model lightweighting, dynamic memory integration multi-voice expansion. Keywords: three-dimensional sequence; convolutional network; folk music; stylisation.
Abstract: This study examines the incorporation of artificial intelligence into intelligent institutions, focusing on the lack of a cohesive closed-loop architecture that integrates multimodal data intelligence, decision analytics, and policy optimisation. The AI-orchestrated digital transformation framework (AI-ODTF) has three levels: data intelligence, decision intelligence, and governance optimisation. multimodal data, relational embeddings, and transformer-based temporal encoding are used to make a single institutional knowledge graph. Ensemble prediction and Bayesian causal inference make it possible to evaluate policies that do not happen, while reinforcement learning with multi-objective optimisation finds the best balance between quality, cost, and sustainability. The framework is confirmed utilising the xAPI-Edu-Data dataset, employing an 80:20 train-test division and five-fold cross-validation. AUC, RMSE, F1-score, and counterfactual gain are all used to measure performance. The results show big improvements, such as an AUC increase from 0.74 to 0.91, a 19.6% increase in resource efficiency, and a 17.9% increase in overall performance. Keywords: artificial intelligence; digital transformation framework; smart universities; data intelligence layer; decision intelligence layer; governance optimisation layer. |
Open Access
Keep up-to-date
-
Our Blog -
Join us on Bluesky -
Follow us on X -
Visit us on Facebook - Our Newsletter (subscribe for free)
- RSS Feeds
- New issue alerts
- Inderscience is a member of publishing organisations including:
-
-
