International Journal of Business Intelligence and Data Mining (24 papers in press)

Special Issue on: Exploring AI Methods and Applications for Data Mining

• Open Access

  An identification method of enterprise financial information transparency based on blockchain ( Free Full-text Access )

  by Yanyan Cao, Xinli Li 
  Abstract: In order to overcome the problems of low financial information recall rate, low recognition accuracy, and long time in traditional methods, a new identification method of enterprise financial information transparency based on blockchain is proposed. Firstly, utilising the decentralised, tamper proof, and traceable features of blockchain technology, a financial information collection system is constructed from the data layer to the application layer; Secondly, key factors affecting the transparency of corporate financial information are screened through multiple linear regression models; Finally, a DCSVM classifier is introduced to overcome the non-differentiability issue of traditional support vector machines through convolutional smoothing loss functions, achieving efficient and accurate recognition of financial information transparency. The experimental results show that the proposed method has the highest enterprise financial information recall rate of 98.05%, the highest recognition accuracy rate of 98.52%, and the shortest recognition time of only 0.64s, with good application effect.
  Keywords: blockchain; corporate financial information; transparency; recognition; multiple linear regression models; DCSVM.
  DOI: 10.1504/IJBIDM.2026.10078432
   
  

Special Issue on: Exploring AI Methods and Applications for Data Mining Part One

• Open Access

  Study on fault diagnosis and recovery for digital distribution networks: ITOT fusion ( Free Full-text Access )

  by Chunmei Zhang, Xingque Xu, Yongjian Li, Silin Liu, Jianyi Li 
  Abstract: This study proposed a fault diagnosis and recovery method of digital distribution network based on ITOT fusion. Build a digital distribution network operation data acquisition architecture using ITOT fusion technology, and perform PCA dimensionality reduction on the collected data. Input the data into a genetic algorithm optimised wavelet neural network to obtain fault diagnosis results. Build a digital distribution network fault recovery model based on the fault diagnosis results, and solve the fault recovery model using the BPSOGWO algorithm to obtain the optimal fault recovery strategy. In diagnosing faults, the proposed method attains a remarkable average accuracy of up to 97.55%, the average fault recovery rate is 96.88%, and the fault recovery time varies between 0.8 s and 1.9 s.
  Keywords: ITOT fusion; digital distribution network; fault diagnosis; fault recovery; optimised wavelet neural network; BPSOGWO algorithm.
  DOI: 10.1504/IJBIDM.2026.10077297
   
  

• Open Access

  Improved multi-objective quantum genetic algorithm for low-carbon economic dispatch in power grids ( Free Full-text Access )

  by Jinhui Ma, Xincun Shen, Li Zhang, Lu Chen 
  Abstract: To improve the low-carbon economic regulation effect of the power grid, reduce carbon emissions and operating costs, a low-carbon economic regulation method for the power grid based on improved multi-objective quantum genetic algorithm is proposed. Firstly, integrate photovoltaic, wind power, energy storage, and electricity market data, use local outlier factor algorithm to clean the data and fill in missing values. Secondly, establish a multi-objective optimization model that includes power generation costs, carbon emissions, and demand response costs. Finally, an improved quantum genetic algorithm is proposed to enhance solution efficiency through quantum gate updates and intelligent population management, achieving low-carbon economic dispatch of the power grid. The results showed that under the control of the proposed method, the highest average carbon emissions were 0.39 tons of CO?/MWh, the highest cost was only 135000 yuan/MWh, and the average new energy consumption rate reached 91.87%.
  Keywords: low carbon economic regulation; multi-objective quantum genetic algorithm; power generation cost; clean data.
  DOI: 10.1504/IJBIDM.2026.10077375
   
  

• Open Access

  Automatic correction of subjective sentence similarity questions based on text image recognition ( Free Full-text Access )

  by Kaichen Tang, Xuyang Tu 
  Abstract: Subjective questions play a vital role in educational and vocational assessments, yet manual grading presents challenges to both efficiency and fairness. To address these challenges in sentence similarity tasks, this study proposed an automated correction method by leveraging text-image recognition. A hardware module for data acquisition via image capture was employed, and a VGGNet-based model was used for highly accurate text recognition. Building on the recognised text, a novel automatic grading approach was introduced that integrated a T5 pre-trained model with a pointer network within a ‘pre-training + fine-tuning’ paradigm. Experimental results demonstrated the effectiveness of the method, achieving a text recognition accuracy of 98.61%, with a low error rate of 2.87% and a processing time of 1.21 seconds. These findings highlight the potential of the system for reliable and efficient automated assessment.
  Keywords: text image recognition; sentence similarity; subjective questions; automatic correction; VGGNet; T5 pre training model; pointer network.
  DOI: 10.1504/IJBIDM.2026.10077486
   
  

• Open Access

  Research on fuzzy clustering of ideological and political MOOC resources under the background of “internet plus” ( Free Full-text Access )

  by Dandan Liu 
  Abstract: The massive, diverse, and complex structure of ideological and political learning materials on MOOC platforms results in low clustering accuracy and high time consumption. This study introduces a novel approach for categorising digital educational resources for political education within MOOC platforms through advanced computational techniques. The methodology begins with comprehensive data acquisition using web-based harvesting tools, followed by rigorous data refinement procedures to ensure information integrity. Subsequently, an enhanced convolution neural network architecture incorporating document embedding vectors was developed for textual feature representation. The framework integrates kernel-based fuzzy c-means clustering enhanced with bioinspired optimisation techniques, specifically employing echolocation-inspired swarm intelligence as the initialisation mechanism for improved convergence. Empirical validation demonstrated superior performance metrics, achieving classification precision above 95% and retrieval effectiveness surpassing the 93% benchmark.
  Keywords: Internet plus; MOOC; ideological and political learning materials; fuzzy clustering.
  DOI: 10.1504/IJBIDM.2026.10078022
   
  

• Open Access

  Research on personalised recommendation of tourist attractions based on improved collaborative filtering algorithm ( Free Full-text Access )

  by Xiaoyan Wang 
  Abstract: This study proposed a personalised recommendation method for tourist attractions based on improved collaborative filtering algorithm. Firstly, by analysing dimensions such as user interaction frequency and the proportion of mutual friends, a dynamic weight system is established to build a trust network. Secondly, the concepts of path trust threshold and longest propagation distance are introduced to calculate indirect trust values. Finally, in terms of improving the collaborative filtering algorithm, three innovative correction factors, namely common user size, rating weight, and mean deviation, were introduced to optimise the similarity calculation and obtain Top-N results for tourist attraction recommendations. The experimental results show that the accuracy of the method proposed in this paper is consistently maintained at a high level of 0.9, with an average absolute error controlled within the range of 0.115 to 0.124.
  Keywords: improved collaborative filtering algorithm; scenic spot; personalised recommendation; trust model; dynamic weight.
  DOI: 10.1504/IJBIDM.2026.10078026
   
  

• Open Access

  A colour restoration method for multi-texture oil painting images based on colour transfer ( Free Full-text Access )

  by Xiaofeng Liu 
  Abstract: Multi-texture oil painting images face challenges such as complex texture interference and artistic feature distortion in the process of colour restoration, making it difficult to accurately restore colour layers and maintain stroke details. Therefore, colour restoration method for multi-texture oil painting images based on colour transfer is proposed. Firstly, the low dimensional subspace denoising method combining RPCA and BM3D effectively removes image noise while preserving texture features. Secondly, design a colour transfer strategy that integrates brightness remapping and K-means clustering to achieve adaptive transfer of reference image colour features. Finally, a generative adversarial network integrating spatial feature transformation and full attention mechanism is constructed, which combines perceptual loss and adversarial loss to achieve visual perception driven colour restoration. The experimental results show that the proposed method has the lowest structural similarity of 0.93 and the highest index of 0.91, both of which are superior to existing comparison methods.
  Keywords: colour transfer; transfer learning; oil painting images; colour restoration.
  DOI: 10.1504/IJBIDM.2026.10078027
   
  

• Open Access

  Optimisation of nursing human resources allocation in large hospitals based on improved particle swarm optimisation algorithm ( Free Full-text Access )

  by Su Yuan, Wan Qing Liang, Duoduo Yu, Huan Liu, Yingnan Tian 
  Abstract: In this paper, an optimisation configuration model based on an improved particle swarm optimisation algorithm is proposed to address issues such as supply-demand imbalance in the allocation of nursing human resources in large hospitals. Firstly, construct a multi-objective function that includes the overall work efficiency and job matching degree, in order to achieve comprehensive optimisation of human resource efficiency and job adaptation. Secondly, the Metropolis criterion of simulated annealing and the crowding factor of artificial fish swarm are integrated into the particle swarm algorithm to reduce the risk of local optima and enhance global search capability. Finally, using an improved algorithm to solve the multi-objective function, output a resource allocation plan that meets the balance of nursing load and job requirements. The experimental results show that the proposed method consistently maintains excellent performance of over 90% in scheduling coverage and 0.88-0.95 in shift balance index testing.
  Keywords: improved particle swarm algorithm; large hospitals; nursing human resources; optimise configuration.
  DOI: 10.1504/IJBIDM.2026.10078046
   
  

• Open Access

  Automatic classification and mining algorithm for massive big data based on machine learning ( Free Full-text Access )

  by Feng Qin, Qiang Wang 
  Abstract: In order to improve the accuracy of automatic classification mining of massive big data and reduce the memory consumption of classification mining, a machine learning based algorithm for automatic classification mining of massive big data is proposed. Firstly, the local outlier factor algorithm is used for data cleaning, and a generative adversarial network is introduced to fill in missing values; Secondly, utilising convolutional neural networks to extract massive big data features and inputting them into support vector machine algorithms in the field of machine learning for classification mining; Then, the particle swarm optimisation algorithm is used to improve the support vector machine, while introducing dynamic inertia weights and position update constraints to enhance the particle swarm optimisation algorithm; Finally, use the optimised support vector machine algorithm to implement classification mining. The results indicate that the ARI value of the proposed method is above 0.962, the highest F1 score reaches 0.967, and the highest memory peak is 31.7GB.
  Keywords: classification mining; massive big data; generate adversarial networks; support vector machine; SVM; particle swarm optimisation algorithm.
  DOI: 10.1504/IJBIDM.2026.10078045
   
  

Special Issue on: Exploring AI Methods and Applications for Data Mining Part Two

• Open Access

  DNN and BiGRU-based hierarchical attention network for intrusion detection ( Free Full-text Access )

  by Hui Yan, Hupeng Liu, Ping Yu, Xiaoqing Xu, Mingxin Li, Yunxin Long, Hanlin CHEN, Qi Wang, Duo Long 
  Abstract: To tackle the rising sophistication of cyberattacks and the critical need to enhance intrusion detection capabilities, this study presents a hybrid architecture combining a DNN with an attention mechanism and a BiGRU with an attention mechanism, BiGRU with attention, and an MLP classifier. The model integrates dual DNN-Attention and BiGRU-Attention submodules to capture high-dimensional static features and temporal dependencies, followed by feature fusion and classification via MLP. Evaluated on NSL-KDD and UNSW-NB15 datasets, the model achieves validation accuracies of 99.48% and 98.43%, respectively, with stable convergence and low loss. Comparative results demonstrate superior performance in accuracy, robustness, and generalization, confirming its effectiveness for practical cybersecurity applications.
  Keywords: intrusion detection; deep learning; bidirectional gated recurrent unit; attention mechanism; network security.
  DOI: 10.1504/IJBIDM.2026.10079023
   
  

Special Issue on: OA Big Data Industrial Application and Computing Innovation Part One

• Open Access

  An operation status prediction of transformer calibration instrument using PSO-attention-LSTM algorithm ( Free Full-text Access )

  by Quan Wang, Jun Zhang, Xia Han, Cong Qi, Yushu Cheng, Jiliang Fu, Jiachuan Long 
  Abstract: To solve the problems of low sampling completeness, low accuracy, and long time consumption in traditional methods, a new operation status prediction of transformer calibration instrument using PSO-Attention-LSTM algorithm is proposed. Firstly, collect multidimensional temporal data through a remote calibration system; Secondly, the SAT-GAN model based on AEGAN is introduced to clean and repair abnormal data; Finally, the PSO-Attention-LSTM model dynamically assigns weights to each time step in the input sequence through attention mechanism, highlights key state information, and captures long-term dependencies of temporal features with the help of LSTM's gating mechanism, thereby achieving prediction of the operating status of the transformer calibrator. In experiments, the proposed technique attains a peak sampling integrity of 97.45%, accuracy rate of 98.60%, and a maximum time consumption of less than 1.19s, which verifies the engineering practicality of this method in predicting the operating status of the calibration instrument.
  Keywords: PSO-Attention-LSTM algorithm; transformer calibrator; operation status prediction; AEGAN; SAT-GAN model.
  DOI: 10.1504/IJBIDM.2026.10078025
   
  

Regular Issues

• Research on the impact of digital marketing campaign strategies on consumer buying intention  
  by Koteswararao Dondapati, Naga Sushma Allur, Durga Praveen Deevi, Himabindu Chetlapalli, Sharadha Kodadi, Thinagaran Perumal 
  Abstract: Consumer psychology and shopping motivation also stay abreast of changes in technology. To address an immense number of audiences and to understand the purchaser’s behaviour, campaigns for digital marketing are pretty crucial for an organisation. However, the purchasing propensity cannot be precisely measured and portrayed by traditional tools. With this limitation, the study managed to deliver an ML-based consumer buying intention analysis method, based on analysis from consumer data through online advertisements using machine learning algorithms. Technique for Order of Preference by Similarity to Ideal Solution applied will allow ML-CBIAM to have precise all-inclusive understanding of customer habit and preference. Simulated results indicate that ML-CBIAM is superior to the state-of-the-art methods in terms of accuracy and coverage in predicting purchase intent through different campaign techniques. This approach helps firms optimise marketing strategies, increase profits, and strengthen customer relationships.
  Keywords: digital marketing campaign; consumer buying intention; machine learning; consumer analysis.
  DOI: 10.1504/IJBIDM.2026.10077344
   
  
• Cloud-based business intelligence in an E-commerce environment for small and medium-sized enterprises  
  by Akhil Raj Gaius Yallamelli, Vijaykumar Mamidala, Mohanarangan Veerappermal Devarajan, Rama Krishna Mani Kanta Yalla, Thirusubramanian Ganesan, Aceng Sambs 
  Abstract: Business intelligence (BI) addresses comprehensive quality management approaches. The interconnected facilities applicable to BI in many instances became more difficult, expensive, and rigid. Therefore, this paper suggests cloud-based business intelligence in the e-commerce environment (CB-BIE) method is used to find solutions for small and medium-sized enterprises. In the CB-BIE method, five customers are interviewed, and a questionnaire session is conducted for 10 small and medium-sized enterprises (SMEs). Each of these aspects covers a particular field that clients pay great attention to when introducing a cloud BI solution. The results indicate that some of the most significant performance measures are software, seamless network services, sensitive responses to requests for customer service, managing vast volumes of data, and implementing costs. CB-BIE method noticed that industry-specific applications for monthly and quarterly charges and emails or phones to access the company are desired.
  Keywords: business intelligence; BI; cloud computing; E-commerce.
  DOI: 10.1504/IJBIDM.2026.10077472
   
  

Special Issue on: Knowledge Discovery from Big Data to Spur Social Development Part 2

• Deep learning-based football training movement analysis and penalty feedback research  
  by Fubin Dai, Shuai Li, Zhigang Li 
  Abstract: Recent advances in video understanding have enabled referee assistance in football, but reliance on single views or costly VAR systems limits their use in training and lower-tier contexts. In this work, we first propose a multi-view deep learning system for analysing football training movements and generating automated penalty feedback. The system processes four-view video sequences through video encoders, followed by a novel cross-view attention fusion module (CAFM) that adaptively integrates features from different viewpoints. Finally, to address view inconsistency and class imbalance in real-world data, we introduce a view completion strategy with augmentation and apply a class-balanced loss for classification. Experiments conducted on the SoccerNet-MVFoul dataset demonstrate that our method achieves 60.67% accuracy and 46.30% balanced accuracy in foul action classification. For foul severity classification, our approach reaches 53.63% accuracy and 54.14% balanced accuracy. Visualisation of attention weights confirms that the model successfully identifies the most informative viewpoints for each foul instance. These results show that the proposed system is effective and interpretable, offering a promising direction for referee assistance in non-professional football scenarios such as youth academies, amateur clubs, and grassroots training sessions.
  Keywords: computer vision; football; video assistant referee; VAR; video classification.
  DOI: 10.1504/IJBIDM.2026.10077556
   
  
• Sports training fatigue recognition using surface electromyography signals on wearable devices  
  by Jing Li, Wenwen Pan 
  Abstract: How to timely assess the fatigue level of athletes to avoid muscle injury is critical for sports daily training. However, it is impossible to use huge device to monitor muscle fatigue level of athletes during training. This paper designs a lightweight muscle fatigue estimation system using surface electromyography (sEMG) signals to tackle these issues. First, the sEMG signals are collected using wireless sEMG sensors worn by athletes. Then, the collected sEMG signals are transmitted to edge device which integrates real-time sEMG signal processing and the lightweight artificial intelligence model deployment. The former one extracts time domain features, frequency domain features and time-frequency features of sEMG signals. The later one adopts relative margin support vector ordinal regression which is sparse to reflect the ordinal relationship between different fatigue levels. The experimental results show the proposed scheme can reach least mean absolute error and satisfy the computing resource limits of edge nodes.
  Keywords: surface electromyography; sEMG; edge computing; fatigue recognition; ordinal regression; wearable devices.
  DOI: 10.1504/IJBIDM.2026.10077587
   
  
• SportVAE: athletic heart rate anomaly detection via wearable sensors using enhanced variational autoencoders  
  by Peng Liu, Yang Yu 
  Abstract: Traditional HRV analysis struggles to distinguish exercise-induced variations from genuine cardiac irregularities, especially in high-intensity activities. To address these issues, this study proposes SportVAE, an enhanced Variational Autoencoder (VAE) for detecting heart rate anomalies of athletes during physical activities via wearable sensors. The proposed model incorporates temporal attention, a bidirectional LSTM encoder for capturing heart rate dynamics, an adaptive weighting mechanism to balance reconstruction error and KL divergence based on intensity, and domain adaptation layers for generalization across sports. Tested on 10,000+ hours of data from 2,000 athletes, it achieved 93.2% accuracy, 91.5% recall, and a 92.3% F1-score, outperforming existing methods. The model adapts across sports, handles varying intensities, and is efficient enough for wearable devices, contributing to both theory and practice in athletic health monitoring.
  Keywords: variational autoencoder; VAE; athletic heart rate monitoring; anomaly detection; deep learning; temporal attention mechanism; wearable technology.
  DOI: 10.1504/IJBIDM.2026.10077588
   
  
• A stroke lesion segmentation method based on volume-balanced data partitioning and dual-branch ensemble network  
  by Siyu Zhao, Baoqiang Li 
  Abstract: Accurate segmentation of ischemic stroke lesions from MRI is crucial for clinical decision-making, including subtype classification and prognosis assessment. However, the heterogeneous size and appearance of lesions in T1-weighted MRI, along with class imbalance, pose significant challenges. In this study, we propose a dual-branch ensemble framework integrating nnU-Net and nnResU-Net to leverage their complementary strengths in global representation and local detail preservation. Furthermore, we introduce a volume-balanced cross-validation strategy to ensure consistent distribution of lesion sizes across training folds, addressing the imbalance problem at the data level. Experiments on the publicly available ATLAS R2.0 dataset demonstrate the superiority of our method. Our ensemble approach achieves a Dice score of 0.6601, volume difference (VD) of 9188 mm³, lesion-wise F1-score (L-F1) of 0.5349, and a simple lesion count (SLC) error of 4.6735 across five-fold cross-validation. These results outperform state-of-the-art baselines, including U-Net, TransUNet, and SwinUNETR. Qualitative visualisation further confirms that our model produces lesion segmentation results most closely aligned with expert annotations. To further enhance clinical applicability, the framework can be deployed in edge-computing environments, enabling low-latency and resource-efficient lesion segmentation close to the point of care.
  Keywords: ischemic stroke; magnetic resonance imaging; deep learning; medical image analysis; segmentation.
  DOI: 10.1504/IJBIDM.2026.10077882
   
  
• Privacy-preserving legal credit risk assessment framework driven by generative AI with interpretability analysis  
  by Di Teng 
  Abstract: The increasing reliance on credit risk assessment models within legal frameworks necessitates a balance between data utility and privacy preservation. This paper introduces a privacy-preserving legal credit risk assessment framework driven by generative AI, specifically leveraging a privacy-preserving generative adversarial network (ppGAN). The proposed framework aims to generate synthetic financial and legal data that preserves privacy while maintaining high utility for downstream legal credit risk models. The core idea of our approach is to leverage generative adversarial networks (GANs) to synthesise data that mimics real-world patterns without compromising sensitive information. Furthermore, we analyse the interpretability of these models, enabling stakeholders to understand the decision-making process behind risk assessments. The ppGAN architecture is evaluated against several baseline methods, including tGAN, MedGAN, HealthGAN, and DP-GAN, in terms of privacy protection, utility (classification accuracy, F1-score), and interpretability. Experimental results demonstrate that ppGAN achieves superior privacy protection and delivers high-performance credit risk predictions.
  Keywords: privacy-preserving; generative AI; legal credit risk assessment; generative adversarial networks; GAN; interpretability analysis.
  DOI: 10.1504/IJBIDM.2026.10078551
   
  
• AI-driven multi-modal legal dispute question and answer system: legal reasoning and evaluation  
  by Lin Lin 
  Abstract: With the growing integration of artificial intelligence (AI) into legal services, AI-driven legal question answering (QA) systems have garnered significant attention. However, most existing systems focus on legal knowledge retrieval and text generation, often failing to effectively address complex legal disputes that require nuanced reasoning based on core legal terms. To address this gap, we propose a novel AI-driven core keyword legal dispute question answering system (CKLQ), which combines core keyword extraction, legal dispute categorisation, and legal knowledge retrieval for handling a variety of legal disputes. It uses a multi-modal reasoning framework to enhance performance in categorising legal disputes and accurately extracting core keywords from complex legal texts. We evaluate CKLQ using accuracy, F1-score, answer accuracy, and explanation accuracy across three comprehensive legal datasets: LawBench, Legal Case Corpus, and Contract Disputes. Performance is validated against five state-of-the-art baseline models including LegalBERT, LexGLUE, LawGPT, LawBench Evaluation (original), and LegalBERT+RAG.
  Keywords: core keyword extractionl; legal dispute classification; legal reasoning; multi-modal reasoning; legal question answering system.
  DOI: 10.1504/IJBIDM.2026.10078552
   
  
• Real-time cross-lingual speech translation for English teaching via federated transfer learning and multi-modal knowledge distillation  
  by Baoying Sun, Yingwei Liu, Ying Huang 
  Abstract: This paper presents a novel approach for Real-Time Cross Lingual Speech Translation using Federated Transfer Learning and Multi-Modal Knowledge Distillation. Real-time translation between languages is critical for many applications in today’s globalised world, but existing solutions often face significant challenges such as high latency, computational overhead, and data privacy concerns. In this work, we address these challenges by introducing a decentralised learning framework using Federated Transfer Learning to enable privacy-preserving model training across edge devices, and multi-modal knowledge distillation to transfer knowledge from a large, accurate teacher model to smaller, more efficient student models. Our approach not only reduces the latency and computational cost of speech translation systems but also ensures high translation quality. Experimental results demonstrate that our model outperforms traditional methods in terms of translation accuracy, latency, and model size, making it well-suited for deployment on edge devices with limited resources.
  Keywords: real-time cross-lingual speech translation; federated transfer learning; FTL; multi-modal knowledge distillation; edge computing; privacy-preserving machine learning.
  DOI: 10.1504/IJBIDM.2026.10078553
   
  
• Athlete fatigue recognition and performance analysis via multimodal deep learning with cross-modal attention mechanisms  
  by Xue Han, Fujiang Cui, Feng Wang, Wei Wang, Haibo Wang 
  Abstract: In modern sport science, timely detection of athlete fatigue and accurate assessment of performance degradation are critical for enhancing training efficiency, reducing injury risk, and optimising competitive outcomes. This study proposes a multimodal deep learning framework with attention mechanisms that integrates physiological (e.g., ECG/EMG), biomechanical (e.g., IMU motion data) and visual (e.g., body pose or facial cues) modalities for simultaneous fatigue-state classification and continuous performance regression. The designed architecture employs dedicated encoders per modality, followed by a crossmodal attention fusion module and dualtask heads for classification and regression. Experiments conducted on publicly available datasets demonstrate that the proposed method outperforms eight baseline algorithms, achieving 95.76% accuracy in fatigue recognition and 0.108 MAE in performance prediction. Ablation studies confirm that each modality and the attention fusion contribute significantly to overall performance.
  Keywords: athlete fatigue recognition; multimodal deep learning; attention mechanism; performance regression; cross-modal fusion.
  DOI: 10.1504/IJBIDM.2026.10078554
   
  
• Intelligent precision advertising via deep learning and multi-source data fusion  
  by Lijing Xu 
  Abstract: With the rapid growth of digital marketing, intelligent advertising recommendation systems play a crucial role in improving delivery efficiency and user experience. Yet, conventional methods often suffer from limited multi-source data fusion, weak interpretability, and poor adaptability across domains. To address these challenges, we propose an intelligent precision advertising framework that integrates heterogeneous data sources using graph neural networks, employs a hybrid deep learning architecture with reinforcement learning for accurate user-ad matching and dynamic optimisation, and incorporates probabilistic modelling to enhance targeting accuracy and cost efficiency. Experiments on real-world datasets show that our approach improves click-through rate by 32.7%, conversion rate by 28.4%, and reduces cost per conversion by 21.5%, while maintaining strong adaptability in dynamic scenarios. These results demonstrate the frameworks potential to shift advertising recommendation from static, rule-based delivery to a personalised, interpretable, and real-time paradigm suitable for next-generation marketing systems.
  Keywords: precise advertising recommendation; artificial intelligence; multi-source data fusion; dynamic optimisation; deep learning.
  DOI: 10.1504/IJBIDM.2026.10078555
   
  
• Real-time lightweight pose estimation for sports motion analysis on mobile and edge platforms  
  by Haibo Wang, Ningning Li, Chao Liu, Bin Wu 
  Abstract: This paper proposes a lightweight motion posture recognition method tailored for real-time sports motion analysis on mobile and edge platforms, which adopts MobileNetV2 as the backbone, augmented with depth-wise separable convolutions and depth-wise transpose convolutions to ensure high-resolution feature up-sampling. A skip concatenation mechanism is introduced to preserve fine-grained spatial information, thereby improving the accuracy of athlete keypoint localisation. To enable real-time feedback in sports scenarios, a full-dataflow pipelining strategy is designed to optimise concurrent operations across feature extraction, pose estimation, and post-processing. Loop unrolling and double buffering are applied to maximise parallelism and minimise memory latency. The proposed approach outperforms state-of-the-art lightweight models on the COCO Keypoints and MPII Human Pose benchmarks. Additionally, it delivers real-time processing speeds of 68 FPS on mobile devices and 180 FPS on edge accelerators. With a compact model size of only 5.2 MB, it ensures efficient deployment in resource-constrained environments.
  Keywords: lightweight pose estimation; mobile motion recognition; real-time processing; depth-wise separable convolutions; DwTConv.
  DOI: 10.1504/IJBIDM.2026.10078556
   
  
• Efficient lightweight AI-driven animation generation for IoT systems via dynamic encoding and contrastive learning  
  by Peng Guo 
  Abstract: We introduce efficient lightweight AI-driven animation generation, named ELA-AG, a novel framework tailored for IoT systems demanding real-time performance under stringent resource constraints. ELA-AG innovates by combining three core components: 1) dynamic encoding to compress temporal-spatial redundancies across adjacent frames; 2) contrastive animation learning to enforce latent alignment across time and eliminate flicker; 3) a resource-aware optimisation objective that integrates model efficiency metrics (FLOPs, memory footprint, energy, latency) directly into training. We evaluate ELA-AG across three datasets — AnimeRun, CartoonMotion-50K and our IoT-AnimeCam — with comparisons against CartoonGAN, AnimeGAN2, DualAST and other baselines. Quantitative results demonstrate that ELA-AG achieves superior perceptual quality (PSNR, SSIM, FID) and temporal coherence (LPIPS, TWE), while achieving notable improvements in efficiency — higher FPS, halved energy consumption, and reduced model size. Comprehensive ablations confirm that ELA-AG achieves a 47% reduction in energy consumption, 1.5 x higher FPS, and a 19.2% lower FID compared with DualAST, while preserving superior temporal coherence. These results set a new benchmark for Pareto-efficient animation synthesis on resource-constrained IoT platforms, ensuring high-quality, temporally consistent animations without sacrificing speed or energy.
  Keywords: efficient animation generation; lightweight AI; dynamic encoding; contrastive animation learning; resource-aware optimisation; temporal coherence; energy-aware AI.
  DOI: 10.1504/IJBIDM.2026.10078607