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 Articles will not be published until the final proofs are validated by their authors.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 (54 papers in press) Regular Issues
Abstract: This paper proposes a model integrating YOLO V11 with the high-resolution network (HRNet) for real-time detection and error correction of pipa fingering actions. The object detection method YOLO V11 is employed to rapidly and accurately locate the left and right hands playing the pipa in video footage. The extracted positioning data is then cropped and fed into the pose estimation algorithm HRNet. By calculating whether the output finger positions deviate from standard angles and coordinates beyond specified thresholds, the model identifies errors such as finger bending or wrist collapse. Through training, the proposed model achieves an average accuracy of 85% at an intersection over union (IoU) threshold of 0.75 for the dual-hand playing detection task. For the hand pose estimation task, it attains an average accuracy of 88% at a target keypoint similarity threshold of 0.75. Keywords: deep learning; real-time object detection; hand gesture recognition; computer vision; high-resolution network; HRNet; intersection over union; IoU. DOI: 10.1504/IJICT.2026.10076597
Abstract: The rapid expansion of the sharing economy has exposed critical limitations in carbon management systems, particularly regarding data transparency and resource circulation efficiency. This study systematically develops an integrated framework that combines blockchain architecture with a dynamic multi-regional input-output model, enabling precise carbon footprint tracking while establishing effective incentives for sustainable practices. The proposed system incorporates an innovative carbon-attributed consensus mechanism that aligns network operations with environmental objectives, alongside smart contract-driven optimisation that enhances circular material flows. Comprehensive evaluation through multi-sector validation demonstrates substantial improvements: approximately 93% reduction in carbon tracking variance, 15%19% decrease in system-level emissions, and blockchain energy consumption reduced to merely 18% of conventional systems. This research provides a verifiable, scalable solution that effectively bridges technological innovation with sustainability goals, offering practical value for stakeholders across sharing economy ecosystems. Keywords: blockchain; sharing economy; proof of equity carbon performance mechanism; circular network; dynamic multi-regional input-output. DOI: 10.1504/IJICT.2026.10076626
Abstract: This study proposes a method specifically designed for recognising students dance teaching movements based on spatial temporal graph convolutional networks. The method employing three parallel spatial temporal aggregation graph convolutions to extract dynamic features. By incorporating OpenPose for human pose estimation, the proposed spatial temporal graph convolutional network-based action recognition method effectively classifies student dance teaching movements. Experimental model achieving recognition accuracies of 89.1% and 90.4% on cross-subject and cross-set benchmarks respectively, compared with existing advanced models, the improvement ranges from 0.3 to 4.0 percentage points and 0.3 to 3.6 percentage points respectively. Furthermore, evaluation of the self-constructed dataset encompassing ten classical dance categories demonstrates that the method achieves recognition accuracy ranging from 89.6% to 96.1%, effectively resolving the issue of inaccurate dance movement recognition. Keywords: spatial temporal graph convolution network; dance movement recognition; human pose estimation. DOI: 10.1504/IJICT.2026.10076627
Abstract: In response to the issue where current English vocabulary pronunciation evaluation models cannot fully extract feature information from different dimensions of spectrograms, this paper first designs a multi-dimensional audio feature extraction algorithm based on multi-scale dilated convolution. This algorithm initially constructs a shallow feature refinement module that uses parallel convolutions to capture time, frequency, and time-frequency three-dimensional shallow features of Mel-frequency cepstral coefficients features. It combines Res2net structure, dilated convolution, and channel attention to capture more fine-grained multi-scale information from the shallow multi-dimensional features. Then it employs a global feature fusion module combined with multiplicative gating mechanisms to enhance cross-scale feature fusion. Finally, differential evolution algorithm optimised support vector machines are used to score the multi-dimensional features. Experimental results indicate that the average evaluation accuracy of the proposed model reaches 94.57%, outperforming comparative models and achieving an objective and accurate assessment of English vocabulary pronunciation. Keywords: multi-dimensional features; machine learning; dilated convolution; support vector machine; SVM; English vocabulary pronunciation evaluation. DOI: 10.1504/IJICT.2026.10076629
Abstract: Traditional manual scoring of English essays faces subjectivity and low efficiency, while existing automated essay scoring (AES) systems suffer from limited genre adaptability, weak scoring-feedback linkage, and poor robustness. To address these challenges, this study proposes a BERT and Zhipu AI model-based automatic scoring algorithm for different English genre essays (BERT-AI-ASEG), integrating overall scoring feature scoring feedback output. Furthermore, an English article intelligent scoring algorithm integrating multiple large language models (EAIS-IMLLM) is developed to enhance robustness. Experiments on the ASAP++ dataset show that the Spearmans rank correlation coefficient and quadratic weighted kappa reach 0.87 and 0.85, improving by 4.82% and 6.25% over comparison algorithms. The proposed algorithm also achieves an evaluative feedback correlation of 0.89 and a multi-genre adaptation index of 0.91 at 150 iterations. These results indicate that the method is effective in improving accuracy, reliability, and teaching feedback effect for large-scale English writing assessments. Keywords: large language model; English articles; scoring; BERT; article genre; Zhipu AI. DOI: 10.1504/IJICT.2026.10076631
Abstract: This study explores the integration of generative AI into creative education and music composition through the regulated use of LSTM-based model generation in the classroom. It demonstrates how AI can enhance creativity while maintaining academic integrity by combining data-driven approaches with ethical safeguards. Findings indicate that generative AI serves effectively as both a compositional aid and a tool for fostering innovative teaching practices. AI has transformed creative domains, offering new opportunities in music-making, yet its educational application requires a balance between innovation and ethical responsibility. Prior research highlights AIs value for both novice and experienced composers, while recent studies emphasise its growing impact on higher education. The research methodology involved data collection, preprocessing, and the development of an LSTM model, alongside safeguards ensuring responsible implementation. Achieving over 87% precision in sequence prediction, the model showed strong performance, and under controlled conditions, students displayed greater initiative and originality. Keywords: generative artificial intelligence; music composition assistance; creative music teaching; long short-term memory; LSTM; AI in education; music pedagogy. DOI: 10.1504/IJICT.2026.10076662
Abstract: To address the critical challenges of limited labelled data and high annotation costs in automatic pronunciation assessment, this study proposes a novel framework that integrates generative adversarial network-based synthetic data augmentation with active learning. A conditional generative adversarial network is employed to generate synthetic speech samples with controlled phonemic and articulatory features, while a hybrid active learning strategy combining uncertainty and diversity criteria is designed to select informative samples for expert annotation. Evaluation on the L2-ARCTIC dataset demonstrates that the proposed approach achieves a Pearson correlation coefficient of 0.843, outperforming the best baseline (0.801) by 5.2%. It also reduces root mean square error and mean absolute error by 8.0% and 9.8%, respectively. The results highlight the synergistic effect of generative and selective data strategies, offering an effective solution for low-resource pronunciation scoring. Keywords: automatic pronunciation scoring; adversarial network generation; active learning; data enhancement; English pronunciation. DOI: 10.1504/IJICT.2026.10076663
Abstract: Abstract pattern design needs breadth without chaos and repeat safety, yet current generators blur structure and drift in colour. This paper introduces a hybrid generator that separates structure from surface using a latent scaffold refined by diffusion and guided by differentiable priors. In the scheme, first a variational encoder learns a compact scaffold for repetition, alignment, and palette allocation; then a latent denoiser sharpens edges and adds micro variation; finally lightweight priors keep symmetry on beat, close borders for seamless repeats, and hold the palette steady with simple knobs. Experiments on two curated corpora show structural coherence rises by 11%, seam energy drops by 51%, and palette error falls by 26% against strong baselines, while sampling stays about half a second at 512 pixels and human preference reaches 71%. Keywords: abstract art generation; diffusion model; variational autoencoder; controllable generation. DOI: 10.1504/IJICT.2026.10076664
Abstract: In this paper, a graph neural network-based cognitive diagnostic model of course association is proposed to address the problem of modelling the association between course knowledge and competency goals in OBE. The model constructs a knowledge topology graph using course prior relationships, aggregates the information of neighbouring course nodes through graph neural networks to enhance the knowledge representation, and combines with the DINA model to achieve fine-grained diagnosis of learners cognitive status. Experiments on the publicly available dataset MoocRadar show that the model in this paper achieves 86.7% in cognitive diagnosis ACC, which is 3.2% and 2.8% higher than the traditional IRT model and the plain DINA model, respectively, and the MSE is reduced to 0.103. The results show that the proposed model can effectively capture inter-course dependencies and support personalised teaching path recommendation under the OBE system. Keywords: outcome-based education; OBE; curriculum linkage; graph neural networks; cognitive diagnostic models. DOI: 10.1504/IJICT.2026.10076665
Abstract: This paper proposes a hybrid communication system integrating low-power wireless mesh networks (LP-WMN), 5G backhaul links, and fibre optic communication to enhance the communication capabilities of remote power infrastructure. The system design includes a LoRa-mesh-based perception layer, a 5G millimetre-wave backhaul layer, and fibre fusion technology at the core layer to overcome the bandwidth bottleneck of the LoRa protocol, ensuring efficient and stable large data transmission. The proposed three-layer joint optimisation mechanism, comprising load-aware AODV routing, power adaptation algorithms, and time-slot scheduling, dynamically adjusts network paths, transmission power, and time slots, effectively improving system throughput, reducing delays, and optimising energy efficiency. Simulation and field test results demonstrate significant improvements in throughput, delay, and energy efficiency compared to traditional LP-WMN systems, with enhanced system stability and reliability. This system provides a reliable communication foundation for future IoT applications, such as smart grids. Keywords: low-power wireless mesh networks; LP-WMN; 5G backhaul; fibre optic communication; hybrid communication systems. DOI: 10.1504/IJICT.2026.10076666
Abstract: With the rapid development of network information technology and internet technology, more and more people choose online learning as an important channel to acquire knowledge. However, while online education is flourishing, users are facing problems such as inaccurate recommendations and imbalanced paths caused by information overload and data silos. Therefore, this article proposes the Federated Union Genetic Recommendation (FUGR) model, which aggregates cross school data using longitudinal federated learning and evolves multi-objective weights using association rule-based genetic algorithms to achieve accurate recommendation of learning resources required by users. This article conducted a series of experiments on the proposed model, and the results showed that FUGR improved the baseline AUC by 7.5%, HR by 6.4%, path consistency by 22%, scarce course coverage by 15%, while maintaining a privacy budget of < 1. Keywords: online education recommendation; federated learning; genetic algorithm; association rules; privacy protection. DOI: 10.1504/IJICT.2026.10076667
Abstract: A semantic parsing model for English video scenes based on sentence semantics and adaptive feature selection is proposed, which dynamically optimises visual features through text guidance and achieves cross modal semantic alignment. The experiment shows that the model performs well on multiple datasets: the average intersection to union ratio on ActiveNet reaches 91, exceeding other models by more than 11 points; Only 60 rounds of training on InternVid achieved an average intersection to union ratio of 90. The model parameter size is 185M, and the inference speed on RTX 3090 reaches 28.5 fps. After lightweighting, the parameter count decreased to 92M and the speed increased to 45.8 fps. This study effectively alleviates the cross modal semantic gap and feature redundancy issues, providing an efficient and accurate solution for intelligent video analysis. Keywords: sentence semantics; adaptive feature selection; video scene parsing; convolutional neural network; CNN; grey wolf optimisation; GWO. DOI: 10.1504/IJICT.2026.10076675
Abstract: This study proposes a learning situation analysis and curriculum planning model that integrates apriori algorithm and heterogeneous graph neural network to address the shortcomings of traditional methods in capturing long-term spatiotemporal dependencies, high-dimensional sparse features, and dynamic adaptability of campus data. The experiment shows that this method achieves an accuracy of 98.26% +- 0.15 in course recommendation, significantly better than comparative models such as collaborative filtering, matrix factorisation, factorisation machine, graph convolutional network, and reinforcement learning (p < 0.001). The rationality score of its course planning reached 9.62, the long-term learning effect improved by 26.95%, and the standardised cumulative benefit index was the highest (0.927 +- 0.006). The research results have achieved collaborative optimisation of behaviour pattern discovery and course semantic reasoning, achieving a good balance between accuracy and interpretability, and providing effective support for intelligent education decision-making. Keywords: behaviour analysis; curriculum planning; innovative curriculum; apriori; HGNN. DOI: 10.1504/IJICT.2026.10076717
Abstract: The traditional swimming analysis method has the disadvantages of strong subjectivity, difficult quantification, high cost, easy to interfere with training, and the existing three-dimensional convolution network is affected by water ripples and bubbles in the underwater scene, and the key feature extraction is insufficient. A swimming motion recognition algorithm based on improved three-dimensional convolution and attention residual network is proposed. The experimental results show that, compared with the baseline C3d model, the number of parameters of the algorithm is reduced by 42%, the floating-point operation is reduced by 58%, the reasoning speed is reduced by 61%, and the recognition accuracy is improved by 8.3% compared with the baseline C3d model under the same hardware configuration. This method provides a feasible and efficient new way for the quantification, real-time feedback and intelligent teaching of swimming technology. Keywords: swimming; feature fusion; improved C3D; channel-spatiotemporal attention; residual network. DOI: 10.1504/IJICT.2026.10076718
Abstract: This study addresses the management challenges of million-kilowatt photovoltaic (PV) stations by proposing an integrated intelligent management and control system. The system combines IoT sensing, edge computing, and AI to achieve dynamic collaborative management. Deployed in a 1 GW demonstration station, the system integrates 256 edge computing nodes covering 128,000 PV arrays. Results show that the average conversion efficiency increased by 2.3%, fault recognition accuracy reached 98.6%, and equipment abnormality response time shortened to 30 seconds. Additionally, combining weather forecasts with grid load data increased annual power generation by 5.8%. Security testing confirms that core function availability exceeds 98% under extreme conditions. The research validates that this cloud edge collaborative architecture significantly enhances the reliability, economy, and operational efficiency of large-scale PV plants, offering a replicable path for the industry. Keywords: million-kilowatt photovoltaic station; intelligent control system; cloud-edge collaborative architecture; power generation efficiency optimisation; troubleshooting response. DOI: 10.1504/IJICT.2026.10076719
Abstract: Social media sentiment analysis faces challenges such as scarce annotated data and complex contextual factors, making it difficult for traditional methods to accurately capture users genuine emotions. This paper proposes an innovative framework integrating transfer learning and semantic enhancement. By pre-training models on massive unlabeled data and combining them with sophisticated contextual semantic enhancement techniques, it enhances the ability to recognise nuanced emotional expressions in English social media texts. Experiments on public datasets like semantic evaluation-2018 convincingly demonstrate that our method achieves an accuracy of 73.12%, representing a 15.32% improvement over traditional approaches. Statistical tests confirm this enhancement is statistically significant (p < 0.01). This research provides more reliable technical support for practical applications such as real-time public opinion monitoring and personalised recommendation systems. Keywords: transfer learning; semantic augmentation; social media; sentiment analysis. DOI: 10.1504/IJICT.2026.10076737
Abstract: To address the challenges of motion uncertainty, difficulty in modelling temporal dependencies, and poor scene consistency in multi-element animation generation, this paper proposes a multi-element animation generation model based on memory-enhanced self-supervised learning and a mixture of density networks. The model first employs a dual-channel convolutional neural network to extract features from animation video frames and linguistic signal data, respectively, and fuses these representations through self-supervised learning tasks. It then utilises a memory matrix to store and recall long-term contextual information, yielding enhanced features. These enhanced features are then fed into a MDN to obtain the probability distribution of the data. Finally, upsampling reconstructs the animation video frames. Comparisons with video diffusion models, StyleInV, AnimateDiff, and DaGAN demonstrate improvements of 6.6, 3.8, 13.1, and 2.3 in FID score, respectively, indicating MEM-MDNs superior capability in generating multi-element animations. Keywords: self-supervised learning; mixture of density network; multi-element animation generation. DOI: 10.1504/IJICT.2026.10076738
Abstract: The rapid growth of online education platforms has led to fragmented data silos in teaching quality metrics, making it challenging for traditional evaluation methods to achieve cross-platform dynamic tracking and analysis while protecting data privacy. This paper proposes an innovative evaluation system based on federated contrastive learning. By introducing a federated learning framework to establish a distributed collaborative training mechanism, it extracts common features from cross-platform teaching data through contrastive learning without sharing raw data. Experimental validation on the public educational network dataset demonstrates that this system elevates teaching quality assessment accuracy to 94.2%, representing a 12.8% improvement over traditional methods, while enabling real-time tracking and dynamic feedback on teaching effectiveness. This research provides an innovative technical pathway and effective solution to the challenge of reconciling data privacy protection with enhanced evaluation efficacy. Keywords: federated learning; contrastive learning; teaching quality assessment; cross-platform analysis; dynamic evaluation. DOI: 10.1504/IJICT.2026.10076739
Abstract: High-dimensional optimisation of partial differential equations often involves large-scale variables and complex constraints, leading to convergence issues during iterative solution processes. To address this, this paper first discretises the original equation using the finite element method to construct an approximate optimisation model, then proposes an improved alternating direction method of multipliers for efficient solution. To further enhance accuracy, second-order schemes and compact difference schemes are employed for temporal and spatial discretisation, respectively, with rigorous convergence proofs established for the discretisation formats. Addressing diffusion-wave phenomena in the problem, a fractional-order physical information neural network is applied to investigate its direct and inverse problems, thereby providing high-quality predicted solutions for the high-dimensional optimisation. Experimental results indicate that the suggested approach converges in just 5.37 seconds, significantly outperforming comparison methods and validating its superior convergence performance for high-dimensional optimisation problems. Keywords: partial differential equation; high-dimensional optimisation; iterative scheme; convergence criteria; finite element method. DOI: 10.1504/IJICT.2026.10076740
Abstract: To address the challenges faced by traditional high jump posture prediction methods - namely environmental noise and motion continuity - this paper proposes an innovative solution integrating diffusion models with neural differential equations. By leveraging diffusion models to learn the true posture distribution from noisy data and employing neural differential equations to capture continuous motion dynamics, the approach effectively overcomes the limitations of conventional methods in modelling complex movements. Experiments on the human3.6m and dynamically informed pose from inertial measurement unit public datasets demonstrate that this method achieves a prediction accuracy of 96.2%, surpassing mainstream long short-term memory models by 8.7%, while reducing joint angle error to 3.2 degrees. These findings provide more reliable technical support for sports science training and sports injury prevention. Keywords: high jump posture prediction; diffusion model; neural differential equation; motion analysis. DOI: 10.1504/IJICT.2026.10076741
Abstract: As smart grid initiatives gain momentum, higher demands have emerged for real-time, precise sensing of power equipment operational status. Traditional live-line detection methods predominantly rely on single sensors, suffering from issues such as low detection efficiency and poor real-time performance. To address these challenges, this paper first designs an edge computing system tailored for live-line data detection in smart grids. Through multi-sensor data fusion, edge nodes can form a collaborative network capable of more rapidly identifying the location of live-line faults. Then, a dual-channel convolutional neural network is designed to accommodate multi-sensor fusion data inputs. A time-frequency attention feature fusion module is incorporated to optimise the feature extraction process. Experimental results demonstrate that the proposed system achieves live detection within 6.3 milliseconds, with a detection accuracy rate of 94.61%. This system enables timely and precise identification of potential equipment faults while maintaining high real-time performance. Keywords: carbon fibre core conductor; dynamic stress field tracking; multi-head sparse attention mechanism; temporal convolutional network; transformer model. DOI: 10.1504/IJICT.2026.10076742
Abstract: This study tackles the crucial problem of forecasting employment anxiety in graduates through surpassing the constraints of traditional evaluation techniques that depend on static self-reports and ignore the dynamic social context. We present a multimodal dynamic graph neural network framework which fuses Smartphone-sensed behavioural data and real-world physical interaction networks. We use a bidirectional gated recurrent unit to encode the temporal behaviour, and we also have a dynamic graph attention to capture the changes in social support. Comprehensive evaluation on the StudentLife dataset shows that our approach outperforms state-of-the-art baselines significantly by reducing root mean square error by 12.5% from 5.82 to 5.09 and mean absolute error by 15.8% from 4.68 to 3.94 and getting a determination coefficient of 0.67. These results show that the framework works well at copying how individual behaviours and social forces work together for mental health checks. Keywords: employment anxiety prediction; multimodal learning; graph neural networks; dynamic social networks; digital phenotyping. DOI: 10.1504/IJICT.2026.10076743
Abstract: In the current era of rapidly developing digital education, single-modal data is insufficient to accurately characterise student learning behaviour. Cross-platform data contains rich information but faces challenges such as data silos and modal heterogeneity. To address this, this paper first collects multi-modal learning behaviour data from students, employs deep learning models to extract text and visual features, and uses attention to fuse multi-modal data. The multimodal fused data is then carefully segmented and distributed to different platform stakeholders. These stakeholders interact and aggregate information through a federated learning framework to obtain more accurate representations of learning behaviour features. Based on this, a Softmax classifier is used to precisely classify cross-platform student learning behaviour profiles. Experimental results show that the proposed method achieves an MAE of only 0.0631, significantly improving the classification accuracy of learning behaviour profiles. Keywords: cross-platform learning; behavioural profiling; multimodal data fusion; federated learning; deep learning. DOI: 10.1504/IJICT.2026.10076744
Abstract: The digital archives of South Koreas intangible cultural heritage contain multimodal resources-images, texts, and audio-posing challenges in cross-modal retrieval due to semantic complexity and the trade-off between accuracy and efficiency. This paper proposes a novel deep hashing network to address these issues. The model employs modality-specific encoders to extract features and a unified hashing layer to generate compact binary codes. A joint loss function is introduced to preserve cross-modal similarity while enabling effective quantisation, enhancing both discrimination and retrieval performance. Comprehensive evaluations on public datasets show that our method achieves a mean average precision of 92.7%, outperforming state-of-the-art approaches by 5.2%, while maintaining real-time retrieval speed. The framework offers a scalable solution that significantly improves accessibility and management for digital cultural heritage platforms. Keywords: deep hash network; DHN; cross-modal search; intangible cultural heritage; ICH; multimedia content; Korean culture. DOI: 10.1504/IJICT.2026.10076745
Abstract: In order to deal with high frequency errors in terms, grammar, and formatting in German software localisation, an automatic detection model is proposed, which integrates error classification and semantics. The model is constructed with 15,379 real error instances, which are combined with a bidirectional long short-term memory with conditional random field (BiLSTM-CRF) semantic recognition pathway. Combined with a customised grammar verification module and abnormal pattern statistic algorithm, a multi-channel cooperative recognition mechanism is formed. Experimental results show that the F1-score is 89.2% across UI strings, interactive prompts, and help documentation scenarios - an 11.3% improvement over single-rule engines and a 3.4% reduction in the false positive rate. The model achieves more than 90% consistency with human review results, showing superior localisation accuracy and processing stability when dealing with complex grammatical structures. Keywords: German software localisation; dependency parsing; localisation quality assurance; linguistic rule modelling; context-aware error detection; CAT tool integration; hybrid detection model. DOI: 10.1504/IJICT.2026.10076746
Abstract: In view of the defects of the existing writing assistance software in identifying errors such as literary texts, to address the deficiencies of existing writing assistance software and improvement model based on fuzzy logic and cognitive load theory. This model constructs a computable cognitive writing load quantification framework and designs a multi-layer fuzzy reasoning system to handle the uncertainty of errors, and then adopts a dynamic programming strategy to generate efficient optimisation suggestions. Experiments based on the public dataset Building Educational Applications 2019 show that this method achieves a score of 62.3% on F0.5, significantly outperforming frontier baseline models such as grammatical error correction: tag not rewrite and text-to-text transfer transformer (p < 0.01), especially showing obvious advantages in dealing with inconsistent styles and logical coherence errors, verifying its effectiveness and advancement. Keywords: fuzzy logic; cognitive load theory; CLT; identification of writing errors; text optimisation; intelligent writing assistance. DOI: 10.1504/IJICT.2026.10076760
Abstract: This paper proposes a meta-learning-based convex optimisation framework to solve the problems of slow convergence, sensitive initial solution and environment mismatch in antenna pattern reconstruction in oral English test system. The proposed framework combines a physically constrained convex model with a lightweight meta-learner pre-trained on electromagnetic simulation tasks to achieve millisecond high-precision beam reconstruction. Experiments show that the proposed method can still approach the optimal reconstruction accuracy at low sampling rate, and the reconstruction speed is more than 5 times faster than the traditional sparse recovery method, and it shows excellent robustness under hardware perturbation and scene switching. This scheme is suitable for resource-constrained sensor nodes, and provides technical support for the application of intelligent reflecting surface and terahertz communication in oral examination environment. Keywords: meta-learning optimisation; antenna pattern reconstruction; convex constraint embedding; wireless sensor networks; embedded lightweighting. DOI: 10.1504/IJICT.2026.10076761
Abstract: Traditional methods for assessing human body aesthetics largely overlook the influence of dynamic three-dimensional postures, limiting their real-world applicability. To address this gap, we propose vision transformer-three-dimensional pose, a novel framework that integrates a vision transformer with graph convolutional networks to achieve accurate pose-aware aesthetic quantification. Our approach begins by extracting global visual features via a vision transformer and concurrently models skeletal kinematics using a graph convolutional network. These cross-modal features are then fused through a gated attention mechanism, guided by biomechanical constraints. Finally, a joint loss function co-optimises pose accuracy and aesthetic perception. Experimental results demonstrate that our method achieves a pose error of 43.7 mm on the Human3.6M dataset and a spearman rank correlation of 0.917 on the South China University of Technology Facial Beauty Prediction 5,500 dataset, outperforming existing state-of-the-art methods and significantly enhancing robustness under diverse viewing angles. Keywords: body aesthetics assessment; vision transformer; cross-modal fusion; joint optimisation. DOI: 10.1504/IJICT.2026.10076762
Abstract: To address the technical challenge of low identification accuracy for crime scene traces in complex environments, this paper proposes a multi-branch feature fusion model based on transformers. Traditional methods rely on manually selected features, while existing deep learning models struggle to balance trace-level details with global structural patterns. To clarify this point, these global structural patterns indeed often encompass critical spatial distribution characteristics of trace features. The relative positions, densities, and geometric relationships of features are fundamental to achieving accurate and robust trace matching. This study constructs four specialised branches - fingerprints, tool marks, footprints, and biological traces - to extract distinctive features from each trace type. It further designs cross-branch interaction and adaptive fusion mechanisms to achieve effective integration of multi-source trace information. Experiments on public datasets demonstrate that our method achieves an average identification accuracy of 95.7%, surpassing mainstream models by over 6%. This significantly enhances trace recognition capabilities in complex scenarios involving blurred or fragmented evidence. Keywords: crime scene traces; transformer model; multi-branch feature fusion; trace identification. DOI: 10.1504/IJICT.2026.10076763
Abstract: To overcome limitations of traditional adolescent health literacy assessment - single data source, insufficient accuracy and interpretability - a multi-source data fusion model is developed. This study has four key phases: first, extract and combine features from four data types (questionnaire, physiological sensor, behaviour record, environment); second, propose a deep learning model integrating a two-path feature encoder and random forest to enhance fitting ability; third, use SHAP and LIME for global and local interpretability analysis to assess key feature weight changes across groups; finally, verify model performance through multiple empirical experiments. Experiments show the integrated model achieves an F1-score of 0.873 and test-set AUC of 0.921, outperforming comparative models significantly while maintaining stable performance across age and gender heterogeneous groups. This paper verifies the validity of multiple data fusion strategies and provides an extensible algorithm framework with an interpretable mechanism. Keywords: health literacy assessment; multi-source data fusion; deep learning; feature importance analysis; cohort analysis. DOI: 10.1504/IJICT.2026.10076764
Abstract: To solve the problems of multi-parameters interdependences and non-linear responses, BP neural network was used. In the framework of traditional BP neural network, we introduce some constraints, such as driving factors, adjustable and regulatable learning coefficients, to improve the convergence and generalisation ability of high dimensional data. Through the establishment of three evaluation measures and the integration of image output module, the model is able to accurately regulate the evolution of density at the same time. The model was trained on a dataset of 960 points. Compared with the conventional BP model, the average UCS prediction error at 28 days is decreased from 0.31 MPa to 0.15 MPa and the total residual value is 13.5%, and the capability of image assessment is increased to 0.902. This demonstrates excellent responsiveness and temporal consistency of the indicators. This model provides technical support and methodological pathways for dynamic quality assessment of airport subgrade construction. Keywords: airport subgrade; enhanced BP neural network; curing effect evaluation; prediction error; visual analysis; momentum optimisation; nonlinear mapping; multi-indicator assessment. DOI: 10.1504/IJICT.2026.10076765
Abstract: This study proposes a cross-prompt English composition automatic scoring method integrating CNN, LSTM, and attention mechanisms (AMs) to enhance generalisation and accuracy. CNN extracts local vocabulary and syntactic features, LSTM captures long-distance semantic dependencies, and the attention mechanism focuses on key scoring information. The final score is generated through a fully connected layer. Experimental results show a Pearson correlation of 0.94 with human scores and an MAE of 0.26, indicating strong correlation and low error. In cross-prompt environmental theme evaluation, the MAE and RMSE were 0.34 and 0.40, with a throughput of 82.11 articles per second and a response delay of 12.09 ms. The proposed method demonstrates high accuracy, efficiency, and adaptability, providing effective technical support for practical automatic English composition scoring. Keywords: automatic grading of English compositions; convolutional neural network; CNN; long short-term memory network; LSTM; attention mechanism; deep learning. DOI: 10.1504/IJICT.2026.10076766
Abstract: This study optimises college vocal teaching by using a neural-network-driven system to address weak personalisation and delayed feedback. Tracking 50 students for 12 months (3,600 hours of audio and behavioural data), we built a model that analyses performance and generates individualised plans in real time. In a three-month intervention, students improved pitch accuracy by 18%, timbre richness by 21%, rhythm control by 24% and emotional expression by 27%. Long-term evaluation showed 35% higher sustained learning motivation than a control group, and overall satisfaction rose to 9.1/10 (+40%). Teachers also reported higher feedback effectiveness, rising from 65% to 88%. Results indicate that neural-network-based personalisation delivers measurable gains in technique, expressivity, and persistence while enhancing instructional efficiency. Keywords: neural network optimisation; personalised teaching strategy; vocal music teaching; data-driven education; intelligent feedback system. DOI: 10.1504/IJICT.2026.10076767
Abstract: Risk forecasting in financial markets is essential in management and engineering choices. In this paper, a more refined XGBoost model that uses big data analytics to identify nonlinear patterns that are usually not detected by the conventional technique is developed. The process can be optimised with a high feature selection, hyperparameter optimisation, as well as advanced preprocessing of the data. The model has been tested and proven to have better performance than the current methods in the market and helps in efficient risk management, resource deployment, and timely risk mitigation. The model is responsive to market changes and is ethical on AI operations in real-time. Combined with live data feeds, it will be able to respond more effectively and promptly to changes, a key that can serve as a powerful financial risk prediction instrument, driven by AI. Keywords: financial market risk; big data analytics; XGBoost optimisation; intelligent decision support system; IDSS; risk management. DOI: 10.1504/IJICT.2026.10076768
Abstract: Traditional translation methods have the deficiencies in context modelling and semantic expression. To address these issues, this study constructs a hybrid model integrating bidirectional long short-term memory (Bi-LSTM), transformer structure, and bidirectional encoder representations from transformers (BERT). We enhance semantic representation and generation capabilities through attention mechanisms and context optimisation strategies. Experimental results show that the proposed model significantly outperforms comparative methods on multiple mainstream evaluation metrics. The model performs prominently in long-distance dependency processing, polysemy discrimination, and complex sentence translation, being able to more accurately maintain sentence structure and improve the naturalness and consistency of translations. The hybrid model integrating deep semantic representation and attention mechanism can effectively improve the quality of machine translation. The research results provide a new technical path for the processing of complex language tasks and are of great significance for the further development of intelligent translation systems. Keywords: machine translation; deep learning; bidirectional long short-term memory; Bi-LSTM; transformer; contextual information. DOI: 10.1504/IJICT.2026.10076784
Abstract: When generating symbolic music, people sometimes encounter problems such as unclear structure or mismatch between tracks. Accordingly, this study proposes a multi-track symbolic music generation method to solve the problems of melodic fragmentation, uneven rhythm, and multi-track conflicts in traditional music generation. Experiments on several music datasets show that compared with other benchmark music generation models, the generated music has a pitch class histogram similarity of 0.91 and a pitch distance of 1.28. The expert evaluation report shows that the structural consistency score is 4.52, and the melody fluency score is 4.56. The synchronisation rate of multi-track rhythm reaches 92%. These findings show that the introduction of structural constraints and cross-track cooperation is effective and can generate more logical and coherent music. This study verifies the feasibility of diffusion model in the generation of symbolic music, and provides a practical multi-track composition method for researchers in other music-related fields. Keywords: symbolic music generation; diffusion model; music structure constraints; multi-track generation; cross-track attention; melody and harmony coordination. DOI: 10.1504/IJICT.2026.10076785
Abstract: At present, natural language generation models have made certain progress, but existing methods still face challenges in maintaining logical consistency, integrating specific domain knowledge, and adapting to cross-domain scenarios. Hence, this paper proposes an optimisation framework that integrates knowledge graph and deep generation model. By introducing rich semantic information from the knowledge graph in the model, the logical consistency, knowledge and linguistic diversity of the generated content are enhanced. This paper conducts an experimental comparison of four models: knowledge-enhanced bidirectional encoder representation from transformers (K-BERT), graph-to-sequence model (Graph2Seq), Coxs proportional hazards model (ST-CoKE), and an optimised model. The experimental results show that the proposed optimised model outperforms the comparison model in several indicators, especially in contextual understanding, knowledge coverage and language diversity. This paper made up for the deficiency of current natural language generation technology in knowledge integration and provided new ideas for solving semantic fault problems. Keywords: knowledge graph; natural language generation; NLG; deep learning; semantic understanding; semantic generation technology. DOI: 10.1504/IJICT.2026.10076786
Abstract: This paper addresses the challenges of accurate axial force control and extended docking time during rapid docking between six-axis robotic arm-mounted hydrogen refuelling guns and hydrogen ports. A multi-scale collaborative perception-driven hydrogen port recognition method and a force/position hybrid control strategy based on task space decoupling are proposed. In the recognition stage, multi-scale collaborative perception is adopted: first, the YOLOv10 algorithm performs coarse-grained image detection to quickly locate candidate regions of the hydrogen port; then, the original image is adaptively cropped based on candidate regions to limit the semantic segmentation scope to the target region, enabling fine-grained recognition. Keywords: multi-scale collaborative perception; hydrogen refuelling port positioning; force/position hybrid control; task space decoupling; model predictive control; MPC; robotic arm hydrogen refuelling docking. DOI: 10.1504/IJICT.2026.10076787
Abstract: Traditional manual methods are often subjective and inefficient, falling short of the demands for large-scale and high-efficiency testing. To address this, an intelligent and automated fitness testing platform has been developed. The study integrates an improved Fast-OpenPose model with a support vector machine for sit-up detection, and combines BlazePose with an enhanced DeepLabV3+ for pull-up evaluation. Finally, a comprehensive platform incorporating both Fast-OpenPose and BlazePose is established. Results indicated that the improved Fast-OpenPose model achieved an average accuracy of 71.2%, an 8.9% increase over the original OpenPose, with a processing speed of 20 FPS. After 2,500 iterations, the refined DeepLabV3+ model attained 78.5% mIoU and 81.7% FWIoU. The platform showed minimal deviation between automated and manual counts for sit-ups and pull-ups, achieving an overall accuracy of 99.82%. This study offers a more intelligent, efficient, and accurate solution for physical fitness assessment, advancing the field toward intelligent automation. Keywords: OpenPose; BlazePose; physical health test; PHT; automation; DeepLabV3+; support vector machine; SVM. DOI: 10.1504/IJICT.2026.10076788
Abstract: Automated code translation is vital for software modernisation. While large language models (LLMs) show promise, existing methods often use uniform strategies for all code complexities and lack specialised error handling. We present AdaptTrans, a method featuring complexity-aware strategy selection and iterative error refinement. It first assesses code translation task complexity: simple tasks receive direct translation, while complex tasks employ specific enhancement strategies. An error-driven iterative refinement system then identifies error types and applies targeted optimisations. Evaluated on CodeNet and TransCoder benchmarks across multiple programming languages, AdaptTrans demonstrates significant improvements in translation accuracy over state-of-the-art baselines. The complexity-aware mechanism reduced average iteration count for simple translations by 8.5% while improving pass rate by 4.1%, confirming enhanced efficiency and effectiveness. Keywords: code translation; large language models; LLMs; automated programming; prompt engineering. DOI: 10.1504/IJICT.2026.10076789
Abstract: In order to explore strategies for enhancing the information technology education capabilities of university teachers, this study proposal is an overall model based on the knowledge framework of technical teaching content, and improvements. Experimental results show that the PA dimension scored the highest (M = 3.95) and the TPCA dimension scored the lowest (M = 1.70), and other dimensions scored between 2.80 and 3.50. A comparison before and after the implementation of the promotion strategy revealed the most significant improvement among teachers aged 41 to 50, with an increase of 53.90%. The proportion of teachers over 50 years old also increased from 8.89% to 40.63%. The strategies proposed in this study have reference value for research on improving various aspects of teacher capabilities, can promote teacher development, and play a role in driving educational industry reform. Keywords: TPACK; informatisation; education capacity; APNN; DDAE-SVR; BPNN. DOI: 10.1504/IJICT.2026.10076854
Abstract: With the growing demand for precise cross-dialect syntactic analysis, this work proposes an end-to-end framework that automatically labels English syntactic variants and quantifies their distribution across dialects. The approach integrates parser-generated silver annotations, a human-audited gold subset, and a dual-head neural model combining CRF-based sequence tagging and span classification. Domain adaptation with gradient reversal, moment matching, and supervised contrastive learning enhances robustness to dialectal shift, while probability calibration ensures accurate rate estimation. Evaluations on multi-source corpora covering American, British, Australian, and Indian English show that the proposed model improves out-of-dialect macro-F1 by 6.9 points over a strong RoBERTa baseline, reduces domain divergence in encoder space by over 55%, and recovers stable, interpretable contrasts for variants such as that-complementiser drop, particle movement, and dative alternation. Keywords: syntactic variation; automatic labelling; cross-dialect analysis; CRF; contrastive learning. DOI: 10.1504/IJICT.2026.10076855
Abstract: To analyse emotions and predict dissemination trends of Sichuan handicraft intangible cultural heritage (ICH) inheritors, this study proposes an ALBERT-based sentiment analysis model and a TCN-based prediction model. Integrating multi-dimensional feature extraction, attention mechanisms, and a dual-channel TCN structure, the models capture long-term dependencies by fusing emotional features with dissemination indicators. Experiments on social media data show the improved ALBERT model achieves 86.21% precision and 83.97% recall in sentiment analysis, outperforming BERT, RoBERTa, and LSTM while reducing memory usage. The TCN-based prediction model attains MAE of 0.389%, MSE of 0.007%, and 91.36% fit, improving accuracy and stability. Emotional distributions exhibit periodic fluctuations during festivals, with dissemination trends driven by both emotional and behavioural indicators. This research enhances ICH emotion analysis and prediction technologies, revealing group emotion patterns and communication evolution to provide data support for government short-video timing decisions and inheritors' content optimisation. Keywords: sentiment analysis; ALBERT; attention mechanism; temporal convolution network; TCN; dissemination trend. DOI: 10.1504/IJICT.2026.10076856
Abstract: Amid globalisation and growing cross-language information needs, machine translation is crucial for overcoming language barriers. Deep learning has advanced it, but transformer faces limitations: insufficient efficiency in capturing long-range dependencies and poor performance in low-resource translation. To address these, this study proposes three core solutions: 1) a hybrid LSTM-transformer architecture fusing LSTMs gating mechanism (long-sequence modelling) and transformers self-attention (global context capture); 2) an adaptive gradient clipping (AGC) strategy for training stability; 3) dynamic weight sharing with adversarial domain adaptation to enhance cross-language transfer. Experiments on WMT14 English-German/French corpora show the models BLEU value is 2.8 higher than benchmark Transformer, with 18% faster convergence; in English Romanian low-resource scenarios, the transfer mechanism boosts BLEU by 5.3. This study validates the hybrid architecture and optimisation strategies, offering new ideas for efficient gradient optimisation and low-resource translation models. Keywords: machine translation; LSTM-transformer; gradient cropping; cross-language transfer; adversarial learning. DOI: 10.1504/IJICT.2026.10076857
Abstract: This study aims to use advanced machine learning models and big data analytics (BDA) to investigate online marketplace client loyalty. To capture nonlinear correlations and identify significant loyalty determinants, the suggested framework uses algorithms like LSTM, XGBoost, support vector machines, and random forests, rather than typical statistical techniques. The study delves into the elements that influence consumer engagement, the frequency of purchases, and the likelihood of repurchases within the fresh food platform and pet-related e-commerce sectors. A ten-phase framework - covering foundation, data collection, preprocessing, AI development, pilot study, validation, and evaluation - is employed to ensure methodological rigour. When compared to more traditional methods of consumer loyalty prediction, LSTM performs better on measures including accuracy, precision, recall, F1 score, and area under the curve (AUC). The findings highlight the role of personalised recommendations, delivery services, and mobile engagement in shaping loyalty, while offering practical strategies for sustainable growth in vertical e-commerce markets. Keywords: e-commerce; consumer loyalty; big data analytics; BDA; machine learning; LSTM; AI marketing; recommendation systems; vertical e-commerce; customer retention; predictive modelling. DOI: 10.1504/IJICT.2026.10076519
Abstract: Delivering temperature-sensitive goods to remote mountain regions is difficult due to sparse infrastructure and challenging terrain. Unmanned aerial vehicles (UAVs) offer a practical option for fast, last-mile transport in such areas, but ensuring cold-chain reliability during flight remains a problem. We present a UAV-based cold chain logistics framework that integrates a hybrid path-planning approach, combining genetic algorithms with simulated annealing to optimise routes for both energy efficiency and terrain constraints. Onboard IoT sensors continuously record temperature and humidity, allowing real-time intervention if cargo conditions drift from required ranges. Tests with four real-world datasets showed up to a 14% reduction in energy use, faster delivery times, and improved temperature stability compared to baseline planners. These results suggest that combining intelligent routing with in-flight environmental monitoring can make UAV cold-chain delivery more reliable in difficult environments. Keywords: UAV logistics; cold chain delivery; genetic algorithm; simulated annealing; internet of things; IoT. DOI: 10.1504/IJICT.2026.10076563
Abstract: Generative artificial intelligence technology provides new insights for art image inpainting and semantic reconstruction. To address the problem of semantically incorrect restoration content in existing research, this paper first optimises generative adversarial network by combining gated convolution with spectral normalisation. Based on this, an image inpainting and semantic reconstruction method is built by integrating text and art image features. The text disentanglement module of the suggested method can obtain key textual features that help restoration. A cross-modal attention module is designed to ensure restored results are as consistent as possible with text semantics. A dual-channel reconstruction module is also designed to enhance the network's ability to predict image structure and text semantics. Experimental results show that the fréchet inception distance (FID) of the proposed method is 3.01, which can restore realistic art images satisfying textual semantics. Keywords: image restoration; semantic reconstruction; generative adversarial networks; multimodal features; cross-modal attention. DOI: 10.1504/IJICT.2026.10076562
Abstract: Global energy demand surge and worsening environmental issues make optimising corporate energy management key to boosting efficiency, cutting costs and achieving sustainability. This study proposes a multi-level, modular decision support system (DSS) architecture for enterprise energy management optimisation, based on big data analysis. It integrates deep learning, reinforcement learning and digital twins, using genetic algorithms (GA) for global search (e.g., multi-energy allocation) and particle swarm optimisation (PSO) for faster-convergent local refinement. Lighting systems account for 47% of production auxiliary energy; office devices take 63.9% of administrative consumption. A 500+-device factory saw energy utilisation rise from 82% to 92% via the system. Analysis shows production-linked consumption (52.7% of total) has core equipment contributing 88.3%. Keywords: big data analytics; energy management optimisation; decision support system; DSS; reinforcement learning; digital twin technology; particle swarm optimisation; PSO. DOI: 10.1504/IJICT.2026.10076518
Abstract: Power transformers are critical in smart grids, but accurate fault diagnosis remains challenging due to limitations in conventional dissolved gas analysis and data-intensive deep learning methods. This paper introduces an integrated deep learning framework combining convolutional neural networks with bidirectional long short-term memory via an adaptive multi-head attention mechanism. The model jointly processes dissolved gas analysis profiles and vibration signatures: convolutional layers extract spatial features from gas data, while bidirectional units capture temporal vibration dynamics. The attention mechanism dynamically highlights discriminative features, improving both interpretability and accuracy. Evaluated on the IEC dissolved gas analysis dataset and real vibration records, our method achieves 98.2% diagnostic accuracy, surpassing support vector machines (85.1%) and deep belief networks (90.8%). Additional tests confirm robustness in detecting incipient faults under varied conditions. Attention weights align with known failure mechanisms, offering a reliable, interpretable tool for predictive maintenance. Keywords: power transformers; fault diagnosis; deep learning; dissolved gas analysis; DGA; attention mechanism. DOI: 10.1504/IJICT.2026.10076596
Abstract: Electric vehicles (EVs) are gaining rapid popularity, necessitating the development of advanced and environmentally sustainable charging infrastructures. This study presents an AI-driven optimisation framework for designing a smart charging control device installed alongside distribution transformers. The proposed system leverages artificial intelligence algorithms to predict energy demand, analyse user behaviour and dynamically adjust charging schedules in real time to improve efficiency. By minimising operational costs, reducing grid congestion, and enhancing energy utilisation, the system maximises overall efficiency and sustainability. Integrating reinforcement learning and predictive analytics further enables adaptive responses to the evolving needs of EV users. Additionally, the system promotes the use of renewable energy sources such as solar and wind power to minimise environmental impact. Experimental results confirm the system's effectiveness in stabilising the grid, optimising energy distribution, and lowering consumer charging costs, demonstrating its scalability and eco-friendly potential within existing infrastructure. Keywords: smart charging; artificial intelligence; optimisation; electric vehicles; grid stability; renewable energy. DOI: 10.1504/IJICT.2026.10076517
Abstract: As cross-border data flows continue to expand, protecting people's privacy has become a major barrier to international cooperation. This research provides a privacy-preserving optimisation technique based on federated learning to address it. In this framework, a cross-border dataset is produced first, and then five main modules are designed. Then, a neural network is used as the base model for federated training to train a global model without sending any data outside of the country. Last, experimental evaluations check how well the system works. The results show that the proposed method has an accuracy of 0.892 and a robustness of 0.912, meaning it works well even while processing many types of data and protecting privacy. This is far better than standard methods. This technology offers a safe and effective way to work together on data across borders. Keywords: federated learning; cross-border data flow; privacy protection; multimodal data. DOI: 10.1504/IJICT.2026.10076595
Abstract: To address the issue of insufficient personalised adaptability in dynamic emotional interventions, this paper proposes a value-oriented meta-adaptive reinforcement learning framework. By integrating meta-learning and reinforcement learning, the framework constructs a dual-level learning architecture capable of rapidly adapting to individual emotional dynamics with minimal interactions. The model employs a multi-objective reward function to synergistically optimise intervention effectiveness, user engagement, and safety. Experiments on public datasets such as Emotional Support Conversation dataset and Distress Analysis Interview Corpus - Wizard of Oz demonstrate that our approach achieves an emotional state improvement rate of 0.78 and a user satisfaction score of 0.82. These results represent significant improvements over traditional deep reinforcement learning and meta-learning baseline models, providing an effective computational paradigm for addressing adaptability challenges in personalised psychological interventions. Keywords: meta-adaptive reinforcement learning; value-based learning; affective computing; psychological intervention. DOI: 10.1504/IJICT.2026.10076561
Abstract: This paper proposes a cognition-aware hierarchical reinforcement learning framework to solve the decision-making inefficiency problem caused by ignoring the user's cognitive status in traditional personalised marketing systems. By formalising cognitive load theory as a reward function and constructing an agent that collaborates metacognition and executive layer, we achieve dynamic optimisation of user cognitive experience while accurately recommending. Experiments based on the real user behaviour dataset of Alibaba Mobile show that the proposed model is significantly better than the state-of-the-art baselines such as deep interest evolution network and deep reinforcement learning for news recommendation in key indicators such as normalised discounted cumulative gain at 10 and conversion rate (p < 0.01). Ablation experiments further verify the necessity of cognitive components and hierarchical architecture. Keywords: personalised marketing; cognitive load theory; CLT; hierarchical reinforcement learning; agent; user experience. DOI: 10.1504/IJICT.2026.10076594
Abstract: Procurement budget forecasting for low-value consumables is critical for corporate cost control. Addressing the limitations of traditional statistical methods in handling demand fluctuations, this study proposes a hybrid machine learning model integrating seasonal decomposition with random forest. Validated using public supply chain datasets, this model reduces the mean absolute percentage error of budget forecasts to 12.3%; significantly outperforming the autoregressive integrated moving average model (18.5%) and linear regression methods (16.2%). Experimental results demonstrate that by integrating temporal characteristics of historical procurement data with external influencing factors, the model achieves a coefficient of determination of 0.89 on the test set. The weighted mean absolute percentage error metric is reduced by approximately 35% compared to baseline methods, providing enterprises with a more precise budget forecasting tool for procurement decision-making. Keywords: low-value consumables procurement; budget forecasting; machine learning; random forest model; WMAPE. DOI: 10.1504/IJICT.2026.10076560 |
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:
-
-
