International Journal of Computational Science and Engineering (35 papers in press)

Regular Issues

• Self-supervised learning with split batch repetition strategy for long-tail recognition  
  by Zhangze Liao, Liyan Ma, Xiangfeng Luo, Shaorong Xie 
  Abstract: Deep neural networks cannot be well applied to balance testing when the training data present a long tail distribution. Existing works improve the performance of the model in long tail recognition by changing the model training strategy, data expansion, and model structure optimisation. However, they tend to use supervised approaches when training the model representations, which makes the model difficult to learn the features of the tail classes. In this paper, we use self-supervised representation learning (SSRL) to enhance the model's representations and design a three-branch network to merge SSRL with decoupled learning. Each branch adopts different learning goals to enable the model to learn balanced image features in the long-tail data. In addition, we propose a Split Batch Repetition strategy for long-tailed datasets to improve the model. Our experiments on the Imbalance CIFAR-10, Imbalance CIFAR-100, and ImageNet-LT datasets outperform existing similar methods. The ablation experiments prove that our method performs better on more imbalanced datasets. All experiments demonstrate the effectiveness of incorporating the self-supervised representation learning model and split batch repetition strategy.
  Keywords: long-tail recognition; self-supervised learning; decoupled learning; image classification; deep learning; neural network; computer vision;.
  
  
• Multilingual language classification model for offensive comments categorisation in social media using HAMMC tree search with enhanced optimisation technique  
  by B. Aarthi, Balika J. Chelliah 
  Abstract: The exponential rise of social media platforms has led to a surge in offensive content, highlighting the necessity for effectively detecting and managing such comments. This necessitates precise and advanced online social networks (OSN) categorisation and optimisation methods. This study introduces and assesses a novel technique for automatically categorising texts, supporting over 60 languages, without relying on a pre-annotated data set. The technique employs multilingual methods based on the randomised explicit semantic analysis (ESA) strategy. To combat the inherently multilingual nature of social media content, the paper introduces an innovative classification and optimisation strategy named hybrid adaptive Markov chain Monte Carlo tree search (HAMCMTS) with enhanced eagle Aquila optimiser (EEAO). The study uses three publicly available datasets to identify negative or offensive comments in various languages, offering a comprehensive analysis in this field. The proposed approach holds potential for diverse applications, particularly in multilingual categorisation tasks such as monitoring disaster-related communications on social media to improve visibility and trust. Moreover, it incorporates a sophisticated mechanism to bolster the dependability of its recommendations.
  Keywords: negative or offensive comments; multilingual languages; explicit semantic analysis; ESA; enhanced eagle Aquila optimiser; EEAO.
  DOI: 10.1504/IJCSE.2024.10066586
   
  
• Enhancing multi-view ensemble learning with zig-zag pattern-based feature set partitioning  
  by Aditya Kumar, Jainath Yadav 
  Abstract: This study suggests a novel approach called Zig-Zag Pattern-Based Feature Set Partitioning. The method involves two steps: first, calculating feature correlations using Pearson's coefficient, and second, ranking features based on mean correlation and arranging them in a zig-zag pattern. The zig-zag pattern ensures diverse and balanced feature subsets, improving model generalization and reducing over- fitting. Experimental results on ten high-dimensional datasets show the practical significance of the suggested strategy, which show that it outperforms previous strategies in accuracy and generalization. This approach advances multi-view ensemble learning, offering a practical solution for improving ensemble model performance in complex data analysis tasks.
  Keywords: feature set partitioning; views construction; ensemble learning; zig-zag partitioning; classification; multi-view ensemble learning.
  DOI: 10.1504/IJCSE.2024.10066742
   
  
• Using generative adversarial network for music transformation  
  by Cheng-Han Wu, Yu-Cheng Lin, Pimpa Cheewaprakobkit, Wan-Chin Ting, Timothy K. Shih 
  Abstract: In this study, we propose a generative adversarial network (GAN) framework for music style transfer. Initially, a dataset of traditional Jiangnan songs is pre-processed into two categories: complete compositions and corresponding musical phrases (starting and ending notes), which are then converted into piano-roll images. The CycleGAN model is then used to train these images until the model converges to establish a music style transfer model. The goal is to allow users to input only the starting and ending notes of each measure as a musical phrase, and the system will convert this phrase into complete musical compositions in the Jiangnan style. Then we use a deep learning framework and music expertise for data processing, enhancing the quality and utility of our conversions. At the same time, we have established music style assessment metrics based on the statistical data of the dataset, providing an effective method for evaluating music styles.
  Keywords: music transformation; generative adversarial network; GAN; automatic music generation; music style transfer.
  DOI: 10.1504/IJCSE.2024.10066865
   
  
• GraphBiGRU model for anti-money laundering based on preference-based reinforcement learning via the label filtering loop mechanism  
  by Meng Li, Xinqiao Su, Lu Jia, Rongbo You 
  Abstract: Anti-Money Laundering (AML) in Bitcoin transactions remains challenging since Bitcoin data has a complex graph structure and sequential nature, with many unknown labels and an imbalanced distribution of licit and illicit transactions. To address these challenging issues, we propose a novel reinforcement learning-based GraphBiGRU model via the label filtering loop mechanism to detect illicit transactions in the Bitcoin blockchain. Specifically, we first constructed the GraphBiGRU network to learn the graph structure and temporal information of Bitcoin data. Then, we introduced the label filtering loop mechanism, which encouraged the GraphBiGRU to select reliable pseudo-labeled samples that reduced data noise interference. In addition, we investigated a preference-based reinforcement learning strategy that enabled the GraphBiGRU to better identify illicit transactions, thereby improving performance on imbalanced datasets. Finally, we conducted experiments on the Elliptic dataset, demonstrating that our method achieved state-of-the-art performance, especially with a limited labelled dataset.
  Keywords: anti-money laundering; illicit transactions; GraphBiGRU; label filtering loop mechanism; pseudo-labeled samples; preference-based reinforcement learning; elliptic dataset.
  DOI: 10.1504/IJCSE.2024.10068029
   
  
• Efficient traffic management in the internet of vehicles through enhanced routing and deep learning  
  by Arundhati Sahoo, Asis Kumar Tripathy 
  Abstract: In the Internet of Vehicles (IoV), vehicles are treated as sophisticated smart devices with robust communication systems. IoV uses cellular technology and internet access for vehicle-to-vehicle communication. However, traditional routing algorithms struggle with rapid vehicle movements and varying road conditions, leading to instability and inefficiency, especially in congested traffic. This study proposes a unique approach called the Improved Greedy-Bi directional Long Short-Term Memory (I-GBiLSTM) predictor, which integrates an Improved Greedy Perimeter Stateless Routing Algorithm to enhance link stability within 5G networks by incorporating real-time data on vehicle movements and road conditions and traffic patterns. Additionally, a BiLSTM neural network has been enhanced by incorporating a 1-dimensional Convolutional Autoencoder (1D-CNNAE) and a Temporal Transformer Encoder (TTE) for monitoring and predicting traffic data, enabling unique feature extraction. Experimental results demonstrate that I-GBiLSTM is superior to the other existing protocols, achieving a 99% delivery ratio, 100 routing overhead, 180 ms end-to-end delay, and 98.2% prediction accuracy.
  Keywords: traffic management; internet of vehicles; IoV; routing; deep learning; network traffic prediction.
  DOI: 10.1504/IJCSE.2024.10068349
   
  
• Social network perspective on false information detection in vehicular ad hoc networks: combining spatial inference with historical behaviour analysis  
  by Youke Wu 
  Abstract: Rapid advancements in vehicular ad hoc networks (VANETs), which create a unique social network ecosystem, have heightened concerns over false information dissemination. To safeguard road safety and VANET integrity, sophisticated data analytics are vital to counteract such misinformation effectively. This necessitates an innovative hybrid detection method, merging spatial inference with historical behaviour analysis, enabling precise identification of false reports in both environmental and human-generated incidents. By refining data gathering and computational processes, this approach eases communication loads while evaluating spatial inference models' efficacy across various VANET scales. A historical vehicle behaviour model bolsters detection accuracy significantly. Through nuanced event categorisation and tailored detection strategies, the hybrid model's effectiveness is reinforced. Findings confirm this strategy enhances detection efficiency and precision, fitting VANET social network monitoring needs. Consequently, policy suggestions involve reinforcing data privacy, upgrading communication infrastructure, and instituting specialised regulatory frameworks to bolster VANETs' resilience against false traffic information.
  Keywords: VANET; false information; hybrid detection; security.
  DOI: 10.1504/IJCSE.2024.10068606
   
  
• ADBSCSL: adaptive DBSCAN-SMOTE with cost-sensitive learning to enhance diagnostic accuracy for imbalanced medical datasets  
  by M. Kavitha, M. Kasthuri 
  Abstract: Medical diagnosis is complicated by imbalanced datasets, which biased models cannot distinguish minority class cases like rare diseases. To improve diagnosis accuracy, the research introduces ADBSCSL, which stands for Adaptive DBSCAN-SMOTE with cost-sensitive learning. Adaptive DBSCAN, SMOTE, and cost-sensitive learning handle skewed data well. Adaptive DBSCAN clusters minority class occurrences. It changes parameters to dataset density change. The diversity of the density condition cannot have caused the minority class to misidentify. SMOTE is then applied to these clusters to increase synthetic examples and class balance. It reduces misclassification costs using cost-sensitive learning. This pushes the model toward minority class priority and avoids majority class bias. The approach was evaluated on Brain Stroke, Cerebral Stroke, and Autism Spectrum Disorder datasets. ADBSCSL F1-scores of 91.8% and 90.6% indicate accuracy over 90% on Brain Stroke and Cerebral Stroke datasets. On ASD datasets, it had 100% accuracy, precision, recall, and F1-score. Results show that the ADBSCSL increases classification performance, making it a powerful and efficient tool for medical diagnosis with highly imbalanced datasets.
  Keywords: imbalanced datasets; DBSCAN; SMOTE; cost-sensitive learning; machine learning; diagnostic accuracy; imbalanced medical datasets.
  DOI: 10.1504/IJCSE.2024.10068671
   
  
• A secure consensus mechanism for IoT-based energy internet using post-quantum blockchain  
  by Yousra Angague, Hadil Sahraoui, Chahrazed Benrebbouh, Houssem MANSOURI, Al-Sakib Khan Pathan 
  Abstract: The Energy Internet (EI) is a cutting-edge technology with a vision to integrate diverse energy sources into an efficient and flexible grid However, there are several critical security challenges for its real-life implementation due to its interconnected nature. Again, the advancement of another cutting-edge technology, Quantum computing heightens these concerns as it threatens the traditional cryptographic defenses, making them vulnerable to quantum attacks In this work, we introduce a secure consensus mechanism for IoT-based EI systems by integrating Post-Quantum Cryptography (PQC) and Blockchain technology In our approach, we combine the strengths of two state-of-the-art protocols: a mutual authentication framework and a post-quantum consensus mechanism. We propose two secure consensus algorithms, leveraging the Dilithium and Falcon PQC schemes. Simulation studies, demonstrate that our proposed approach significantly improves transaction throughput and reduce latency, providing a resilient framework for secure energy data management in post-quantum world.
  Keywords: blockchain; consensus; energy internet; EI; internet of things; IoT; post-quantum cryptography; PQC; quantum attack; security.
  DOI: 10.1504/IJCSE.2024.10068855
   
  
• Air pollution prediction by using long-short-term memory neural network  
  by Qinghua Xu, Jiankang Shen, Meng Gao 
  Abstract: High ground-level ozone concentrations affect air quality, plant growth, and human health. This study uses an LSTM model to predict 1-h, 8-h, and 24-h ozone concentrations. We tested models with various hidden layer neurons and sequence lengths. Sensitivity to parameters rose with longer prediction intervals. After optimizing hyperparameters, LSTM outperformed traditional methods like Random Forest and MLP in predicting ozone concentrations, with satisfactory predictive capability and pollution event warning rates.This validates the feasibility of LSTM models for predicting environmental ozone levels across different time intervals and confirms their effective ability to forecast air pollution incidents effectively.
  Keywords: ozone; long-short-term memory; hyperparameters; time interval.
  DOI: 10.1504/IJCSE.2025.10069211
   
  
• VMO-HNIA: virtual machine optimisation using hybrid nature inspired algorithm for cloud resources efficiency  
  by Ruaa Ali 
  Abstract: Optimisation for cloud data centres virtual machine (VM) consolidation is advised, however performance trade-offs are difficult. VM optimisation utilising the hybrid nature inspired algorithm (VMO-HNIA) is a new VM consolidation framework. The HNI-based VM consolidation system uses a multi-resources aware decision algorithm (MADA) to identify host overload or underload dynamically. To enhance the optimisation of VM resources and load balancing, the MADA calculates numerous resources to inform decision-making. The correct classification of each host further boosts VM consolidation processes like VM selection, migration, and placement. To improve the process of selecting and placing VMs in a VM consolidation architecture, we suggest using a new technique called the hybrid whale optimisation technique (HWOA) for VM selection and placement. To improve VM consolidation, the HWOA places the best host utilising several objective functions. Experimental findings show the VMO-NHI framework employing CloudSim outperforms underlying solutions.
  Keywords: cloud computing; decision-making; host placement; nature-inspired; VM selection; VM placement; resources optimisation.
  DOI: 10.1504/IJCSE.2025.10069245
   
  
• Pedestrian head detection based on improved YOLOv5  
  by Yong Ren, Tian Qiu, Jian Shen 
  Abstract: This paper presents an improved YOLOv5 model for the detection of pedestrian heads in crowded scenes. By incorporating FasterNet, the C2f module, Soft- NMS and Optimal Transport Assignment (OTA), the proposed model achieves significant performance improvements over the baseline YOLOv5s model, with a recall of 75.21%, AP50 of 84.31%, and AP50-95 of 57.29%, while maintaining a reduced computational complexity of 14.2 GFLOPs. In comparison with other YOLO series models, the proposed model demonstrates a higher AP50-95 score while maintaining competitive recall and AP50 values. The effectiveness of the model has been demonstrated in diverse scenarios, including various crowd densities, lighting conditions, pedestrian orientations, image resolutions, and pedestrian sizes. The results indicate that the improved YOLOv5 model exhibits robustness, adaptability, and generalization capabilities in challenging pedestrian head detection tasks.
  Keywords: pedestrian head detection; YOLOv5; FasterNet; soft non-maximum suppression; Soft-NMS; optimal transport assignment; OTA.
  DOI: 10.1504/IJCSE.2025.10069246
   
  
• Feature selection using war strategy optimisation algorithm for software fault prediction  
  by Pradeep Kumar Rath, Roshan Samantaray, Susmita Mahato, Sushruta Mishra, Sanat Kumar Patro, Himansu Das 
  Abstract: Identifying problematic software modules early on in development process can help programmers create software that is highly efficient and dependable. In this paper, a novel FS approach using war strategy optimization (FSWSO) have been proposed that applies ancient war strategy planning principles to the selection of features or variables in SFP. This approach seeks to identify the most relevant features for SFP by simulating army operations and evaluating the performance of different feature subsets in a simulated war space. In this experiment, we have compared the proposed FSWSO algorithms performance to that of other FS techniques including FSACO, FSDE, FSGA, and FSPSO in order to assess the algorithms accuracy. In the majority of cases, FSWSO has provided better performance with fewer chosen features. The suggested approach has been validated and proven to be superior to prior approaches in choosing an optimal selection of features using the Friedman and Holm tests.
  Keywords: software fault prediction; war strategy optimization; metaheuristic; machine learning; feature selection; classification.
  DOI: 10.1504/IJCSE.2025.10069977
   
  
• Face spoofing detection using noise-based random feature and Fisher vector encoding  
  by Fang Xu, Na Yang, Xiaochao Zhao, Hao Chen, Manzoor Ahmed, Yi Ma, Zhen Liu, Yuquan Zhang 
  Abstract: With the vast application of face recognition technology, its security risks have increased as systems are vulnerable to spoofing attacks with falsified faces, attracting many researchers attention. In this paper, we proposed to make use of noise information in colour space to detect spoofing attacks. Firstly, we extracted frame-based noise from face videos in multiple colour spaces. Then local random features are extracted via random projection. Finally, Fisher vector encoding is employed to aggregate these features into global feature vectors, and a classification model is trained for detection. Experimental results on three standard face spoofing databases demonstrate the effectiveness of the approach. The equal error rate on the replay attack database is 0%. On the CASIA and MSU databases, the equal error rates are 3.52% and 0%, respectively. By combining noise-based random features and Fisher vector encoding, this method effectively resists photo and video-based spoofing attacks.
  Keywords: face recognition; face spoofing detection; noise; random projection; feature extracting.
  DOI: 10.1504/IJCSE.2025.10070188
   
  
• TDSSO: Tasmanian devil squirrel search optimisation enabled deep learning for ambiguity removal in aspect-based sentiment classification  
  by Neelima S. Ambekar, Anant V. Nimkar 
  Abstract: An aspect-based sentiment classification is a fine-grained approach for extracting relevant information from online customer reviews. Accurate evaluation of such reviews helps to understand customer demands and develop new marketing strategies. The proposed work addresses the problem of eliminating ambiguity about relevant feature extraction using context keywords. A Deep Learning (DL) approach, Hierarchical Deep Learning for Text (HDLTex), is used for sentiment categorization based on the multiple features mined from the review data. Additionally, we employ an optimization algorithm known as the Tasmanian Devil Squirrel Search Optimization (TDSSO) to estimate the weight parameters of HDLTex. The proposed method outperforms the state-of-the-art models in assessing the efficacy of sentiment classification based on k-fold values and training data. Experimental results demonstrate significant improvements in True Positive Rate, True Negative Rate, and Testing Accuracy with values of 0.926, 0.909, and 0.947, respectively.
  Keywords: aspect-based sentiment classification; ambiguity; optimisation; deep learning; context keywords.
  DOI: 10.1504/IJCSE.2025.10070598
   
  
• DHARM: recurrent convolutional neural network for real-time activity recognition with smartphone sensor data  
  by Sourav Bera, Anukampa Behera, Chhabi Rani Panigrahi 
  Abstract: Recently, the field of activity recognition using sensor data has been quickly growing worldwide, leveraging the concepts of deep learning (DL) and artificial intelligence (AI). The proposed Dynamic Human Activity Recognition Model (DHARM), utilizes the power of recurrent convolution neural network using Long-Short Term Memory (RCNN-LSTM) which is tested on two publicly available datasets, as well as the raw real-time data collected. The pre-processed time series data is passed to the RCNN-LSTM for feature extraction. DHARM, has been experimented to show a training accuracy of 96.82% and a test accuracy of 98.32%. The model records a low latency of 43 ms for inference or prediction, which makes it effective for real-time human activity detection. This model is robust, reliable in many locations and environmental situations and may be used for real-time Human Activity Recognition in a wide range of use cases.
  Keywords: dynamic human activity recognition model; DHARM; recurrent convolutional neural network using long-short-term memory; RCNN-LSTM; deep learning; accelerometer; gyroscope.
  DOI: 10.1504/IJCSE.2025.10070638
   
  
• Prediction model for recruitment of railway bureaus and enrolment of railway schools based on deep learning  
  by Haijun Wang, Wei He, Junlun Sun 
  Abstract: With urbanization accelerating, the demand for railway transportation is increasing, making it essential to plan recruitment for railway bureaus and adjust enrollment at railway schools. This study aims to accurately predict recruitment needs using historical data. We applied deep learning models, including Back Propagation Neural Network (BP Neural Network), Long Short-Term Memory (LSTM), and LSTM-Attention, to forecast recruitment numbers for eight positions across 18 railway bureaus in 2025, yielding MAE values of 100000, 0.16, and 0.13, respectively. We also used Linear Regression, Ridge Regression, LASSO Regression, and Random Forests to predict the number of remaining graduates in eight major railway programs for 2025, with most models showing MSE values between 0 and 4. Finally, we established upper and lower limits for vocational student enrollment quotas in 2025 by applying factors of 80% and 75% to the predicted recruitment numbers. These findings provide valuable insights for recruitment and enrollment planning, enhancing the application of deep learning in railway recruitment forecasting.
  Keywords: long short-term memory; LSTM; attention mechanism; ridge regression; LASSO regression; random forest; prediction model.
  DOI: 10.1504/IJCSE.2025.10070809
   
  
• FPrune: a parameter pruning algorithm based on federated deep classification model  
  by Xinjing Li, Zheng Huo, Teng Wang 
  Abstract: Federated learning is a distributed machine learning framework that enables multiple participants to train models collaboratively without sharing raw data. However, significant data transmission is required for parameter communication. As deep neural network models grow in size, deploying federated learning in complex network environments results in substantially increased communication costs. To address this challenge, we propose a pruning algorithm for deep federated text classification models, called FPrune. This algorithm evaluates the importance of locally trained models during the federated learning training stage by calculating the importance of each filter. Filters with lower importance are pruned. Additionally, we introduce a bidirectional pruning strategy that prunes filters on both the client and server sides. Experimental results demonstrate that the FPrune/25% and FPrune/50% algorithms reduce the communication cost by 70.22% and 42.03%, respectively, compared to FedAvg. Furthermore, the model’s performance loss is limited to approximately 1.34%, demonstrating that the FPrune algorithm can effectively reduce communication costs while maintaining minimal performance degradation.
  Keywords: federated learning; deep classification; model pruning; TextCNN.
  DOI: 10.1504/IJCSE.2025.10070810
   
  
• Decision network for interpretable UAV decision making under human supervision  
  by Vinicius Abrão Da Silva Marques, Ana Lidia A. Castro, Braulio M. Horta, Ney R. Moscati, Carlos H. Q. Forster 
  Abstract: With the increasing number of drones flying simultaneously under the supervision of a limited number of humans, it is essential to delegate more decision-making authority to the autopilot system to minimize intervention. The decision making capabilities of autonomous UAVs have a direct impact on mission safety and reliability. In this paper we propose a decision-making model for UAVs based on Decision Networks.The Non-Dominated Sorting Genetic Algorithm (NSGA III) is used to train the model from a provided set of cases. Separate selected cases with the respective expected outcomes were used to test the model, showcasing its capacity to accurately represent cases while maintaining interpretability. We demonstrate that this model can facilitate mission accomplishment with risk minimization under uncertainty in stochastic environments. In particular, the model provides a meaningful interpretation of the parameters and how they are taken into account when making decisions.
  Keywords: decision network; decision making; UAV; optimisation.
  DOI: 10.1504/IJCSE.2025.10070933
   
  
• Chinese text-oriented sentiment analysis models, corpus, and recent advances  
  by Zhongliang Wei, Chang Ge, Chang Su, Jun Zhu, Guangli Zhu 
  Abstract: Sentiment analysis of Chinese text presents unique challenges due to the distinct characteristics of the Chinese language, including cultural nuances, word formation styles, and the dynamic nature of certain terms on social media. This paper reviews recent methods for Chinese sentiment analysis, which can be broadly classified into three categories: dictionary-based, traditional machine learning-based, and advanced deep learning-based. A comparative analysis highlights the strengths and limitations of each method across various applications. Additionally, commonly used Chinese corpora, sentiment analysis systems and tools are introduced. Furthermore, this paper discusses the potential directions for future research, such as recognizing complex sentiment states, multimodal sentiment analysis, and cross-cultural sentiment analysis.
  Keywords: Sentiment analysis; Machine learning; Chinese social networking.
  DOI: 10.1504/IJCSE.2025.10071176
   
  
• Inpainting method of generative adversarial network based on improved reconstructed loss function  
  by Yin E. Zhang, Xiao Wen Ye, Qi Zhou 
  Abstract: In processing image restoration tasks, more and more researchers are using generative adversarial networks, and has made many good achievements. However, the loss function used in current repair methods does not give different weights to the losses of areas with different repair effects, making it difficult to repair areas with poor repair effects. Therefore, we propose a generative adversarial network repair method based on improved reconstructed loss, which assigns different weights to the losses of areas with different repair effects by improving the reconstructed loss, enhancing the repair effect on areas with poor repair effects. We will also introduce improved reconstruction loss into perception loss and style loss to enhance the restoration effect of image details. At the same time, in response to the problems of insufficient feature extraction and low feature utilization in current repair methods, our encoder and decoder use a multi-scale dense convolution module (MDCM) to more fully extract image features and improve feature utilization efficiency. And residual spatial attention module (RSAM) was introduced in the skip connection and decoder.
  Keywords: image inpainting; dilated convolution; generative adversarial network; improved reconstruction loss function.
  DOI: 10.1504/IJCSE.2025.10071304
   
  
• Enhancing fairness in deep learning: key tasks, measurement methods, and experimental validation  
  by Xiaoqian Liu, Weiyu Shi 
  Abstract: Deep learning is an important field in machine learning research. It has powerful feature extraction capabilities and superior performance in numerous applications, including computer vision, natural language processing, and speech recognition etc. However, unfairness in deep learning models has increasingly harmed people's interests. Therefore, designing methods to effectively enhance fairness has become a major trend in the development of deep learning. This work reviews key tasks and fairness measurement methods in deep learning. In addition, we conduct experiments on typical fair deep learning datasets to implement individual fairness. The experimental results show that a balance is achieved between accuracy and fairness of classification tasks.
  Keywords: deep learning; algorithmic bias; individual fairness.
  DOI: 10.1504/IJCSE.2025.10071366
   
  
• Design and implementation of personalised recommendation system for university library based on GNN and data fusion  
  by Jie Yang 
  Abstract: Traditional libraries face challenges such as sparse data, cold start issues, and insufficient personalisation in resource recommendations due to their resource-centric service model. To address these issues, this study developed a personalised recommendation model using natural language processing and graph neural networks. The model integrates multi-dimensional data from university students and faculty at the feature layer and analyses the influence of neighbours in different graphs to predict user preferences more accurately. Experiments on Yelp and library datasets demonstrated that the proposed model outperformed six other recommendation systems, achieving the lowest mean absolute error (MAE) of 0.149 and a stable root mean square error (RMSE) of 0.2451. By leveraging social data to enhance user behaviour analysis, this approach alleviates cold start problems and improves recommendation accuracy. The method also indirectly boosts reader satisfaction, offering practical value for personalised book recommendations in libraries.
  Keywords: data fusion; user portrait; graph neural network; GNN; library; personalisation; recommendation; Word2Vec.
  DOI: 10.1504/IJCSE.2025.10071858
   
  
• Temporal similarity-constraint graph networks for stock prediction with stock relations  
  by Jincheng Hu, Yu Zhang 
  Abstract: Stock prediction aims to enhance investment decisions by forecasting future stock trends, traditionally using time-series data. While deep learning has advanced time-series modeling, most existing methods treat stocks as independent entities, overlooking the rich relationships between them. Additionally, conventional approaches frame stock prediction as a regression problem focused on price prediction, which does not align directly with investment goals. To address these issues, we propose Temporal Similarity-constraint Graph Networks (TSCGN), a novel framework that incorporates stock relations and selects stocks with the highest return ratio. TSCGN embeds sequential stock data into features and constructs a stock knowledge graph to capture interactions between stocks. By integrating temporal similarity constraints, TSCGN enhances prediction accuracy and robustness. Experiments on real-world datasets (NASDAQ and NYSE) demonstrate that TSCGN outperforms state-of-the-art models in prediction accuracy and investment returns, making it a valuable tool for financial decision-making.
  Keywords: graph networks; similarity constraint; stock prediction; stock relation.
  DOI: 10.1504/IJCSE.2025.10072064
   
  
• QGA-optimised BL xLSTM MLP model for portfolio  
  by Meng Li, Zhihui Song, Jiaxu Feng 
  Abstract: The Black-Litterman (BL) model integrates market conditions and investor judgments. However, existing research on generating expert insights focuses on either regressing returns against external variables or modeling return series as time series, without integrating both approaches. Furthermore, traditional BL portfolio optimization neglects transaction costs and fails to optimize hyperparameters, limiting its adaptability to varying market conditions. To address these, we propose a QGA-optimized BL_xLSTM_MLP model, that combines external variable regression (via MLP) and time-series modeling(via xLSTM) ,which integrates temporal dependencies and macroeconomic features into expert views, while optimizing hyperparameters and transaction costs using a Quantum Genetic Algorithm (QGA). The QGA adopts the sum of the Sharpe ratio and Information ratio (accounting for transaction costs) as the fitness function, effectively addressing the traditional BL model’s neglect of transaction costs. Finally, we conducted experiments on USA 30 industry portfolio demonstrating that our method achieved state-of-the-art performance.
  Keywords: Black-Litterman model; quantum genetic algorithm; xLSTM model; asset allocation; portfolio.
  DOI: 10.1504/IJCSE.2025.10072092
   
  
• An irony detection model for social media comments based on topic context  
  by Yanhui Wang, Yuhao Zhou, Zixuan Zhang, Shunxiang Zhang, Lei Chen 
  Abstract: Due to the ever-expanding amount of information available on social media comments, the need for reliable and efficient irony detection mechanisms becomes evident. However, most current methods focus on analysing the comment text, ignoring the role of the topic context in promoting the understanding of ironic semantics. In order to provide a more efficient framework for extracting ironic text features, we propose an irony detection model for social media comments based on topic context (SMC-TC). The model is mainly composed of the information union module and the sentiment-guided module. In the information union module, the topic-comment text features are obtained through Bi-GRU and attention-over-attention (AOA) to realise the combination of topic context and comments. In the sentiment-guided module, the comment word embeddings are input into LSTM and combined with the self-attention mechanism to get the sentimental semantic features. Based on the above two modules, the sentimental semantic features and topic-comment text features are concatenated and then fed into a sigmoid function to acquire the final irony detection result. Through extensive experiments on the publicly available ToSarcasm dataset, our model can enhance the F1-score to 74.04%.
  Keywords: irony detection; social media comments; topic context; Bi-GRU; attention-over-attention.
  DOI: 10.1504/IJCSE.2025.10072093
   
  
• Achieve Sim2Real based on semantic constrained cycle generative adversarial network  
  by Xiangfeng Luo, Hongbin Huo, Xinzhi Wang 
  Abstract: In the field of vision-based control systems, the discrepancy between simulator and real-world environments renders models trained in simulators ineffective in real-world scenarios. Previous approaches have attempted to mitigate this issue by mapping the simulator and realworld into a shared latent space, but this can result in the loss of semantic information relevant to decision-making in the images. In this paper, we propose a method called Semantically Constrained CycleGAN (SCCGan) to address these limitations. SCCGan extracts semantic information from generated images and compares it with the original images to ensure consistency. Experimental results demonstrate that our method preserves the semantic information of the original images during the generation process, enabling the transfer of decision models from simulators to the real world. By leveraging semantic constraints, SCCGan facilitates the effective migration of decision models, bridging the gap between simulated and real-world environments in vision-based control systems.
  Keywords: simulator to reality; CycleGAN; reinforcement learning; autonomous decision making.
  DOI: 10.1504/IJCSE.2025.10072111
   
  
• Hybrid predictive modelling for insurance premium retention: integrating statistical and AI techniques  
  by Ahmed A. Khalil, Zaiming Liu 
  Abstract: This research highlights the critical role of forecasting in the insurance industry and emphasises the premium retention ratio (PRR) as a key internal performance indicator for evaluating insurance company operations. Traditional time series models like ARIMA and exponential smoothing face limitations in capturing complex data patterns. To address this, the study proposes a hybrid predictive model that combines statistical time series models (ARIMA, EXP) with advanced AI techniques (ANN, SVR) to enhance PRR prediction accuracy in Egypt's Fire, Marine, and Aviation insurance sectors. Using 80% of data for training (1989–2015) and 20% for testing (2016–2021), the study demonstrates that hybrid models, particularly ARIMA-ANN and EXP-ANN, outperform conventional models. The findings suggest that incorporating ANN into these models significantly improves prediction accuracy. This research offers a novel approach to forecasting in the Egyptian insurance market and provides publicly accessible datasets for further comparative studies across different countries.
  Keywords: artificial neural network; ANN; artificial intelligence; autoregressive integrated moving average; ARIMA; exponential smooth; Egyptian insurance market; statistical time series; support vector machine; SVM; insurance.
  DOI: 10.1504/IJCSE.2024.10067258
   
  
• A green pattern-based data encryption solution in the cloud  
  by Farah Abdmeziem, Saida Boukhedouma, Mourad Chabane Oussalah 
  Abstract: Organisations are increasingly interested in leveraging cloud computing for extensive data storage. However, security concerns hinder its widespread adoption due to the need to entrust data to cloud providers, which reduces customer control. A viable solution is customer-side data encryption, but this can be resource-intensive and lead to performance issues and higher carbon footprints. To address this, we propose a solution that employs customised encryption/decryption patterns and categorises data into three sensitivity levels, while also accounting for data access and update frequency. Furthermore, we consider cost metrics to assess the environmental implications of encryption. These metrics enable to evaluate our proposed solution while highlighting its positive environmental impact in comparison to other state-of-the-art approaches. The results show that our approach not only facilitates the adoption of encryption mechanisms but also manages the fine balance between fulfilling data confidentiality demands and effectively handling practical resources and energy constraints.
  Keywords: data; security; cloud computing; green encryption; encryption-decryption patterns; cost metrics.
  DOI: 10.1504/IJCSE.2024.10066044
   
  
• Performance analysis and comparison of jelly-fish optimisation-based maximum power point tracking controller for partial shading condition  
  by Dilip Yadav, Nidhi Singh 
  Abstract: This paper addresses the critical challenge of partial shading condition (PSC) in photovoltaic systems, which significantly affect the efficiency of PV panels. Conventional methods often fail to optimise output under partial shading condition, prompting the need for innovative approaches. The study proposes the jelly-fish optimisation algorithm for maximum power point tracking, comparing its effectiveness with various existing MPPT controllers including incremental conductance, modified incremental conductance, perturb and observation, particle swarm optimisation, cuckoo search algorithm, grey wolf optimisation, and whale search optimisation techniques. The study reveals the limitations of conventional techniques in optimising power output under PSC. The findings highlight the superiority of the jelly fish-based MPPT, achieving an impressive efficiency of 99.89% with a minimal tracking time of 0.14 seconds, surpassing other MPPT controllers. This work advances the field by highlighting the jelly-fish algorithm's effectiveness and guiding future research toward more efficient MPPT methods.
  Keywords: cuckoo search; jelly-fish optimisation algorithm; maximum power point tracking; MPPT; partial shading condition; PSC; particle swarm optimisation; PSO.
  DOI: 10.1504/IJCSE.2024.10067209
   
  
• Scalable malicious URL detection technique for smishing attacks  
  by Razvan Stoleriu, Catalin Negru, Bogdan-Costel Mocanu, Emil-Andrei Constantinescu, Alexandra-Elena Mocanu, Florin Pop 
  Abstract: Nowadays, smartphones are used daily and use sensitive information making people more vulnerable to cyber-security attacks. The easiest way for attackers to access a smartphone is through SMS phishing (smishing) using URL shortening services. In this paper, we propose a scalable technique to detect malicious URLs in smishing attacks based on a cloud-edge architecture, using threat intelligence platforms (e.g., VirusTotal and PhishTank), and machine learning algorithms that classify the URLs based on their features. We used a public dataset for training and proposed new features to improve it. We evaluated our proposed ML model against JRip, PART, J48, and random forest algorithms. Our model has improved performance compared to similar solutions, obtaining an accuracy of approximately 97%. To showcase the effectiveness of our solution, we implement an Android application that detects malicious short URLs in SMS messages and notifies the user concerning their legitimacy.
  Keywords: smishing attacks; malicious URLs; edge-cloud computing; threat intelligence; machine learning.
  DOI: 10.1504/IJCSE.2024.10067011
   
  
• EIUAPA: an efficient and imperceptible universal adversarial attack on audio classification models  
  by Huifeng Li, Pengzhou Jia, Weixun Li, Bin Ma, Bo Li, Dexin Wu, Haoran Li 
  Abstract: The domain of audio classification models is emerging as a significant paradigm, albeit susceptible to universal adversarial attacks. These attacks involve the insertion of a single optimal perturbation into all audio samples, leading to incorrect predictions. Nonetheless, existing attack methodologies are hindered by inefficiencies and imperceptibility challenges. In order to streamline the attack process effectively, we propose a two-step strategy EIUAPA that offers an optimal initiation point for the perturbation optimisation process, resulting in a notable decrease in generation time. To maintain imperceptibility, we present a range of metrics focusing on perturbation concealment, serving as benchmarks for optimisation. These metrics ensure that perturbations are not only concealed in the frequency and time domains but also remain statistically indistinguishable. Experimental results demonstrate that our method generates UAPs 87.5% and 86.8% faster than baseline methods, with improved signal-to-noise ratio (SNR) and attack success rate (ASR) scores.
  Keywords: adversarial attack; artificial intelligence; security and privacy; audio classification; deep learning.
  DOI: 10.1504/IJCSE.2024.10066803
   
  
• Constructing stock portfolio with transformer  
  by Jinyuan Li, Linkai Luo 
  Abstract: Machine learning methods have been applied to quantitative investing, yet the application of transformer models remains limited. Stock prices are influenced by both long-term and short-term features. Existing methods usually treat the influencing factors as a whole and do not distinguish them. In this paper, we introduce a transformer encoder-decoder architecture tailored for the capture of long-term and short-term features. By partitioning historical data into long-term and short-term parts, the encoder module concentrates on extracting long-term features, while the decoder concentrates on short-term features and the integration of long and short-term features. Portfolios are then constructed from the top N predicted stocks. Experimental results show that the proposed transformer model outperforms the existing state-ofthe- art methods, LSTM, RNN, and GRU models, with improvements of 26%, 19%, and 14% in annualised returns for long-short portfolio combinations, respectively. It indicates the benefits of extracting long-term and short-term features separately.
  Keywords: stock portfolio; transformer; factor model.
  DOI: 10.1504/IJCSE.2024.10066373
   
  
• MECNet: multi-modal edge co-guidance network for RGB-D salient object detection  
  by Xiuju Gao, Chenxing Xia, Xia Chen, Jianhua Cui 
  Abstract: Currently, mainstream RGB-D salient object detection (SOD) methods rely on depth information to supplement RGB information, which may suffer from inappropriate results due to simplistic fusion. Furthermore, numerous existing methods mainly yield saliency maps with erroneous or blurry edges because of their inadequacy in harnessing edge details and local information from RGB and depth images. Therefore, we propose a multi-modal edge co-guidance network (MECNet) for RGB-D SOD. Firstly, a multimodal attention fusion module (MAFM) is designed to fuse RGB and depth information effectively. Moreover, an edge co-guidance module (ECM) which uses edge consistency between RGB images and depth maps to capture the coherence edge information is developed. In the decoding process, a level-by-level cascade fusion is performed through simple element-wise addition. Finally, the coherence edge information is integrated into the saliency output stage to generate clear and sharp saliency maps. Extensive experiments illustrate the effectiveness and generalisability of our method.
  Keywords: coherence edge; multimodal fusion; RGB and depth images; salient object detection; SOD.
  DOI: 10.1504/IJCSE.2025.10071715
   
  
• A verifiable and secure DNN classification model over encrypted data  
  by Weixun Li, Guanghui Sun, Yajun Wang, Long Yuan, Minghui Gao, Yan Dong, Chen Wang 
  Abstract: Outsourcing deep neural networks (DNN) is beneficial to reduce the client overhead, but there are sensitive data privacy issues. However, the existing schemes not only fall short in privacy-preserving and gradient integrity verification, but also fail to perform complex nonlinear operations. In this paper, we design, implement, and evaluate a verifiable and secure DNN classification model over encrypted data (DNNCM-ED), which provides confidentiality and integrity verification simultaneously. Firstly, we propose a new framework in which the client and model training servers jointly train the DNN model to achieve the model training server for aggregation. Secondly, we design secure communication protocols for basic operations, which can be used to construct DNN classification model. Finally, we further devise a verifiable algorithm related to the DNNCM-ED, which provides confidentiality and integrity verification simultaneously. Extensive property and performance analyses indicate that DNNCM-ED is effective, as well as sharing communication and computation overhead of the cloud.
  Keywords: deep neural networks; encrypted data; homomorphic encryption.
  DOI: 10.1504/IJCSE.2024.10066642