International Journal of Computing Science and Mathematics (13 papers in press)

Regular Issues

• Design of personalised knowledge teaching platform based on C4.5 algorithm and association rules  
  by Shasha Chen 
  Abstract: To address the limitations of traditional teaching models in accommodating individual student differences, this study developed a personalised knowledge instruction platform integrating C4.5 algorithm and association rules. The model first constructs personalised knowledge profiles using C4.5 algorithm to extract learning behaviours and preferences. By combining an enhanced Apriori algorithm with data snapshot and hash index mechanisms, it significantly improves rule generation efficiency and recommendation accuracy. Experimental results demonstrate that the model achieves 98.4% accuracy in FIC dataset and 98.2% in ASSISTments dataset, with F1 scores reaching 91.14%. Key metrics including support, confidence, and gain reach 90.73%, 84.22%, and 93.86% respectively, outperforming competing models like LightGBM and XGBoost. This study validates the platforms effectiveness and practical value in personalised teaching recommendations, providing a scalable technical framework for intelligent education systems.
  Keywords: C4.5 algorithm; association rules; teaching; personalisation; recommendation.
  DOI: 10.1504/IJCSM.2026.10078105
   
  
• Mental health diagnosis model based on multi-task and multi-view learning  
  by Xiaoyu Yang 
  Abstract: Mental health diagnosis remains a global challenge due to limitations in existing models, which often rely on single-task learning or unimodal data sources, leading to low accuracy and poor generalisation across diverse user groups. To address these issues, this paper proposes a novel mental health diagnostic model based on multi-task and multi-view (MTMV) learning. A self-constructed dataset combining smartphone behavioural data and psychological scale responses was preprocessed using a Word2Vec-based Skin gram model, with signal-to-noise ratio (SNR) and information loss rate used to evaluate preprocessing performance. The proposed MTMV framework integrates attention-enhanced feature fusion, expert gating, and cross-task representation sharing to jointly diagnose anxiety, stress, and depression. Experimental results show that the model achieves 97.6% accuracy and 89.2% recall for students, and demonstrates superior performance across multiple occupational groups. These findings indicate that the MTMV-MH model offers a robust and scalable solution for intelligent mental health assessment.
  Keywords: multi-task; multi-view; mental health; Word2Vec; MTMV; SNR.
  DOI: 10.1504/IJCSM.2026.10078231
   
  
• A Study of Film Music Emotion Recognition via Multimodal Fusion Strategy  
  by Li Zhan, Na Wan, Ming Zhao 
  Abstract: Film music constitutes one of the most essential components in cinematic creation. It is widely adopted to shape atmospheric tones, portray characters’ inner psychological states, and reinforce the dramatic emotions conveyed in films. To scientifically verify whether film music can induce varied emotional responses and specific affective categories in audiences, this study extracts valid features from both film music and scripts. Each single-modal input is then imported into its corresponding pre-trained large model to capture modality-specific emotional information. After that, the multi-source emotional outputs are integrated and analyzed to categorize emotional types. Experimental results indicate that the improved multimodal fusion approach presented in this paper significantly boosts the performance and reliability of emotion recognition, benefiting from the complementary characteristics of distinct modalities. Accordingly, this work offers meaningful references for the field of film music emotion recognition.
  Keywords: Emotion recognition; Multimodal; Film music; Film script; Multimodal fusion.
  DOI: 10.1504/IJCSM.2026.10079339
   
  
• Approximate bisimulation of semi-algebraic transition systems based on dynamically weighted metrics  
  by Guxuan Li, Weidong Tang, Meiling Liu 
  Abstract: With the widespread application of embedded systems, program verification faces challenges of state space explosion and non-deterministic branch redundancy. To address the static error control limitations of traditional singular value decomposition (SVD) methods in approximate bisimulation, this paper proposes a dynamically weighted approximate bisimulation (DyWAB). The method integrates Wus characteristic set method with SVD techniques, introducing a reconciliation-based error compensation mechanism derived from the harmonic mean principle. A dynamic weighting allocation strategy is designed based on singular value characteristics to suppress error propagation. Combined with row-minimal matrix determination and multi-candidate selection optimisation, DyWAB achieves system simplification under strict error control. Experimental results demonstrate that, compared to static threshold baselines, DyWAB reduces validation errors by 5177% and improves state approximation accuracy by 22%, validating its effectiveness in synergistic optimisation of error control and system simplification.
  Keywords: program verification; dynamically weighted; SVD; approximate bisimulation; Wu's characteristic set method; multi-candidate selection.
  DOI: 10.1504/IJCSM.2026.10079567
   
  
• Experimental Simulation of Logistics Demand Forecasting Based on Fuzzy Cognitive Map Model and Ant Colony Algorithm  
  by Wenqian Han, Zixia Chen, Yue Diao 
  Abstract: This study has established a logistics demand forecasting model based on Fuzzy Cognitive Map (FCM), optimizes the FCM weight matrix using the Ant Colony Algorithm (ACA), and introduces the Niche Technique for the design of the Ant Colony Algorithm-Fuzzy Cognitive Map (ACA-FCM) based logistics demand forecasting method. To effectively evaluate the application effect of ACA-FCM model on logistics demand forecasting, this paper innovatively proposes a logistics demand forecasting method based on ACA-FCM model, and uses Pandas and Python tools for experimental simulation and analysis of logistics demand forecasting. The results demonstrate that the ACA-FCM model outperforms the GWO-FCM and GA-FCM models in three aspects: comparative testing of convergence performance across different models using both standard and noisy logistics datasets, testing predictive performance by introducing accuracy, precision, and recall metrics, and in terms of stability and predictive accuracy for logistics demand forecasting, achieving the best overall performance.
  Keywords: Fuzzy Cognitive Map (FCM); Ant Colony Algorithm (ACA); Niche Technique; Logistics Demand Forecasting.
  DOI: 10.1504/IJCSM.2026.10079622
   
  
• A calibration transfer of sediment carbon content between a lab spectrometer and a hyperspectral camera using deep learning algorithms without standard samples  
  by Pingping Fan, Yong Wang, Zijian Wang, Xueying Li, Huimin Qiu 
  Abstract: Hyperspectral imaging has become the preferred technology for large-scale and high-frequency ecological environment monitoring due to its advantages in image analysis, but it requires point calibration. Visible and near infrared spectroscopy can be used for real-time calibration of hyperspectral data due to its fast, accurate, and point analysis features. Therefore, it is important to study the calibration transfer between laboratory spectrometers and hyperspectral cameras. This article studied the calibration transfer of sediment carbon content between laboratory spectrometer QE65000 and hyperspectral imaging camera without standard samples, using a deep learning method based on multilayer autoencoder (MAE) and generative adversarial network (GAN) without standard samples. Results showed that both MAE and GAN could achieve good transfer with R2 of 0.885 and 0.792, and RMSE of 1.258% and 1.528%, respectively. These transfer were much better compared with the traditional sample free calibration transfer such as the parameter-free framework calibration enhancement (PFCE). This study explored the sample free calibration transfer between different types of instruments which provided an effective reference for other types of calibration transfer.
  Keywords: sample free calibration transfer; sediment; reflectance spectroscopy; hyperspectral; deep learning.
  DOI: 10.1504/IJCSM.2026.10080019
   
  
• Dynamic Early Warning of Coal Mine Floor Water Inrush Using Cost-Sensitive Random Forest  
  by Jie Lou, Baohui Chen, Xiaoyong Yang, Wenchao Ye, Xiang An, Xin Guo, Weiming Chen 
  Abstract: Addressing class imbalance and dynamic response issues in coal mine floor water inrush evaluation, this paper proposes an early warning method using Game Theory weighting and Cost-Sensitive Random Forest. The mining process is discretized into time slices, constructing a dynamic-static indicator system including factors like distance to working face. Logarithmic transformation normalizes long-tail data, and Game Theory optimizes weights to validate the water inrush coefficient. For prediction, an asymmetric cost matrix and safety threshold optimization enhance Random Forest, minimizing expected risk and improving recognition of Level IV Critical risks under small-sample conditions. Test results show the model suppresses majority class bias, achieving complete recall of Critical risks, thus providing quantitative support for deep mine water hazard management.
  Keywords: Coal mine water hazard; Combination weighting; Random Forest; Safety threshold optimization; Imbalanced learning.
  DOI: 10.1504/IJCSM.2026.10080254
   
  
• Ada-ISO-PINN: Solving Parabolic Partial Differential Equation Using Optimized Hybrid Loss-Based Physics-Informed Neural Network  
  by Ganesh B. Dapke, Govardhan G. Bhuttampalle 
  Abstract: Partial differential equations (PDEs) are utilised to learn all types of common phenomena and to describe various physical phenomena. Several methods have been developed to offer accurate estimations of parabolic PDEs, but they have certain limitations, such as lower performance, higher computational cost, and time complexity. Hence, an adaptive intelligent searching optimisation enabled physics informed neural network (Ada-ISOPINN) model is proposed to solve the parabolic PDEs. The efficiency of the Ada-ISO-PINN model is obtained by incorporating the Ada-ISO algorithm, which aids in resolving the parabolic PDEs accurately with minimal errors in the solutions. The utilisation of hybrid loss functions helps to measure the error value accurately; moreover, it supports the Ada-ISO-PINN model to learn in a more informative and meaningful manner, which helps to minimise the error value. The validation results demonstrate that the Ada-ISO-PINN model attains an RMSE of 2.07, a correlation of 0.93, MAPE of 9.09, MAE of 1.13, R2 of 0.88, and MSE of 4.29, which is far better than the traditional methods.
  Keywords: Deep Learning; Adaptive Intelligent Searching optimization; Parabolic Partial Differential Equations; Hybrid loss function; Physics Informed Neural Network.
  DOI: 10.1504/IJCSM.2026.10080543
   
  
• A Cross-modal Evolution Reasoning Framework for Mountainous Transmission Line Safety via Edge-deployed Vision-Language Models and Dynamic Knowledge Graphs  
  by Qi Cao, Song Yang, Hao Lin Li, Xin Qiang Wu, Ming Liu, Zhijun Qin, Feng Wang 
  Abstract: Mountainous transmission lines face compound hazards, while traditional monitoring suffers from semantic ambiguity and isolated multi-modal data. This paper proposes a cross-modal reasoning framework integrating edge vision-language models (Edge-VLM) and cross-modal dynamic knowledge graphs (CDKG). We design a lightweight Edge-VLM with decoupled contrastive loss and NF4 quantisation, reaching mAP@50=0.94 on Jetson Orin Nano. A semantic-guided binocular reconstruction achieves millimeter-level depth precision complying with power industry standards. CDKG fuses InSAR, meteorology and visual data via R-GCN for causal risk propagation. Tested on Jilin-Mount-24 dataset, the framework delivers a 10-hour advance early warning for geological disasters with full reasoning interpretability.
  Keywords: Transmission Line Inspection; Vision-Language Models; Knowledge Graphs; Edge Computing; Multi-modal Fusion; InSAR.
  DOI: 10.1504/IJCSM.2026.10080881
   
  
• Optimization and Simulation of Pickup and Transportation Scheduling for Digital Freight Platform  
  by Wenwen Zhang, Zixia Chen, Qianqian Wang, Haosheng Zhang, Xiaoping Chen 
  Abstract: This study addresses vehicle dispatching in pickup transportation for digital freight platforms, tackling multi-depot dynamic scheduling complexity and real-time response limitations. A novel KNN-GA two-stage hybrid algorithm is proposed, integrating K-nearest neighbours (KNNs) clustering for spatial decomposition and an improved Genetic Algorithm for route optimisation. The multi-objective model considers dispatching costs, temporal costs, and load factors. KNN preprocessing reduces search space complexity by 83.3%, while elite preservation and dynamic parameter adjustment enhance convergence. Experiments on automotive supply chain data demonstrate convergence to a fitness value of 0.015 within 30 iterations, with 26.5% computation time reduction (180 s vs. 245 s), 85.3% average load factor, and an 8.7% empty-running rate (p < 0.01, Cohens d = 1.82). The algorithm enables minute-level dynamic rescheduling for 50-vehicle fleets at $2.10 per AWS EC2 run, facilitating the transportation management system (TMS) integration via RESTful APIs. This hybrid integration of spatial clustering and evolutionary optimisation fills the research gap in closed-loop simulation for order-dispatch-pickup processes.
  Keywords: Fuzzy Cognitive Map (FCM); Ant Colony Algorithm (ACA); Niche Technique; Logistics Demand Forecasting.
  DOI: 10.1504/IJCSM.2026.10080885
   
  
• Failure Modes and Effects Analysis using Data Envelopment Analysis under Neutrosophic Environments  
  by Reihaneh Hafizi Atabak, Hossein Sayyadi Tooranloo, Mohammad Zarei Mahmoudabadi 
  Abstract: Failure Mode and Effects Analysis (FMEA) is a widely used method to identify defects in the design of a product or process, but it has limitations in handling uncertain or incomplete information. Therefore, this study proposes an input-oriented data envelopment analysis (DEA) model that is combined and improved with netrosophic logic to transform FMEA analysis in environments full of uncertainty. In contrast to the traditional FMEAs reliance on deterministic numerical scores, this model uses netrosophic set theory to simultaneously capture correctness, uncertainty, and incorrectness, allowing for a three-dimensional representation of the uncertainty inherent in expert judgments. Designed to assess the relative efficiency of organizations with comparable input-output structures, this framework extends the scope of FMEA to complex scenarios. In order to demonstrate the effectiveness of the model, a numerical example is introduced to show how the model can handle uncertainty and provides a transformative advance in risk assessment methodology.
  Keywords: Failure Modes and Effects Analysis (FMEA); Data Envelopment Analysis (DEA); Neutrosophic; Risk Management.
  DOI: 10.1504/IJCSM.2026.10080886
   
  
• An Imbalanced Small-Sample Hydropower Unit Shaft Trajectory Recognition Method Based on Improved Adaptive Empirical Wavelet Transform and Convolutional Neural Network  
  by Yingbing Ran, Xihui Zhang, Yujie Zhao, Jiajin Yuan, Kongyun Chen, Jian Wang 
  Abstract: To address limited sample size and class imbalance in shaft center trajectory signals of hydropower units, which degrade the accuracy and generalization of conventional classifiers, this paper proposes an imbalanced small-sample identification method integrating empirical wavelet transform (EWT) and convolutional neural network (CNN). EWT first performs multimodal decomposition to suppress noise and extract frequency-band-specific features. A particle swarm optimization (PSO) algorithm then adaptively refines EWT modal boundaries, generating optimized multi-channel trajectory images. To mitigate imbalance and underfitting, a binary ensemble strategy decomposes the multi-class task into several independent binary classifications. The optimized multi-channel images are subsequently input into a CNN for feature learning and recognition. Experimental results show that the proposed approach outperforms traditional SVM and single-channel CNN models across multiple trajectory identification scenarios. PSO optimization improves overall accuracy by about five percentage points over the baseline, enhancing robustness and reliability.
  Keywords: Hydroelectric unit; axial trajectory; fault diagnosis; unbalanced small sample.
  DOI: 10.1504/IJCSM.2026.10080888
   
  
• Double rough set models for acquiring more accurate knowledge at lower cost  
  by Wanting Wang, Qingzhao Kong 
  Abstract: At present, almost all rough set models are proposed by using one information table. It is difficult to effectively reduce the cost of collecting data labels and accurately acquire knowledge when applying these traditional rough set models to analyze data. To address this dilemma, for any given concept, the samples in the positive and boundary regions obtained from the original information table are considered as the universe of the second information table, and three types of dual rough set models (DRSMs) are established based on different learning tasks. Research shows that DRSMs not only greatly reduce the cost of collecting data labels but also effectively improve the accuracy of acquired knowledge.
  Keywords: rough sets; granular computing; information table.
  DOI: 10.1504/IJCSM.2026.10080891