International Journal of Computational Science and Engineering (8 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;.
  
  
• X-SSAS: human-machine interaction-driven framework for explainable and scalable similarity-based link prediction in social networks  
  by Mridula Dwivedi, Vipin Saxena, Babita Pandey 
  Abstract: This study proposes a novel unified similarity-based link prediction method, X-SSAS, that addresses the limitations posed by traditional methods and serves as a medium to explore human-machine interaction. X-SSAS combines both structural similarity and attribute similarity controlled by a tuning parameter to generate efficient similarity scores that considers both network topology and node attributes. The proposed study provides a novel approach to k-medoid clustering using X-SSAS scores and offers transparent predictions using explainable artificial intelligence. Extensive experiments are conducted on 14 real-world datasets and X-SSAS is tested using six well-known evaluation metrics. The results demonstrate that X-SSAS outperforms existing studies and four widely used similarity measures in their weighted version. The proposed X-SSAS framework attains 99.86% accuracy and 99.99% AUROC on the C. elegans datasets. X-SSAS provides enhanced accuracy, transparency, trustworthiness, and scalability across diverse real-world network datasets.
  Keywords: link prediction; explainable artificial intelligence; XAI; k-medoid clustering; HMI; social network; structural similarity; attribute similarity; Jaccard’s coefficient; Euclidean distance.
  DOI: 10.1504/IJCSE.2026.10077984
   
  
• Hybrid convolutional network with adaptive graph attention mechanism for HSI classification  
  by Sandeep Swamy, Shashidhar Sonnad 
  Abstract: This research introduces a novel hyperspectral image classification (HSIC) model, named 3D-2D hybrid convolution and enhanced graph attention mechanism (HCN-AMGAM), designed to overcome the challenges of high-dimensional spectral-spatial feature extraction and classification accuracy. HSI captures rich spectral information across multiple wavelengths, enabling precise identification of diverse ground objects. The proposed HCN-AMGAM integrates 3D-2D hybrid convolution with an enhanced graph attention mechanism to jointly exploit spatial continuity and spectral correlation. The hybrid convolution module effectively captures local-global dependencies while reducing computational complexity compared to conventional 3D CNNs. An adaptive graph construction strategy based on deep spectral-spatial embeddings strengthen feature connectivity, and the multi-feature fusion GAM enhances information integration while minimising redundancy. This synergy leads to more discriminative and robust hyperspectral representations. The proposed end-to-end framework not only achieves higher accuracy but also enhances model interpretability and scalability for real-world remote sensing tasks. Experiments conducted on Kennedy Space Center (KSC) and Pavia University (PU) datasets demonstrate that HCN-AMGAM achieves state-of-the-art performance, confirming its significant impact on advancing hyperspectral image classification and environmental monitoring applications.
  Keywords: hyperspectral image; HSIs; convolutional neural networks; CNNs; Pavia University; PU; Kennedy Space Center; KSC; graph attention mechanism; GAM.
  DOI: 10.1504/IJCSE.2026.10078381
   
  
• Generating empathetic responses via knowledge augmentation and chain-of-thought in conversational information retrieval systems  
  by Zhinan Gou, Yan Li, Mengyao Jia, Siyu Liu 
  Abstract: Conversational information retrieval systems stand as a crucial research area at the intersection of information retrieval and artificial intelligence. However, existing research largely neglects the generation of emotional responses. This paper proposes a novel framework for generating empathetic responses by integrating knowledge augmentation and chain-of-thought reasoning. Firstly, we leverage the knowledge-augmented capabilities of the large language model to generate semantically similar questions, acting as supplementary data to expand the coverage of dataset and enrich the contextual diversity. Then, relevant top answers are retrieved using the original and semantically similar questions, with semantic similarity from datasets as the key metric. Moreover, we establish a chain of thought that guides the large language models to generate empathetic responses, enabling them to offer targeted suggestions for alleviating emotions associated with questions of users. Experimental results show that our method significantly outperforms the baselines in terms of emotional expression and user-perceived empathy.
  Keywords: empathetic response; knowledge augmentation; chain-of-thought; large language model.
  DOI: 10.1504/IJCSE.2026.10078433
   
  
• A software fault prediction model using driving training-inspired metaheuristic feature selection framework  
  by Himansu Das, Pratiti Mishra, Sanat Kumar Patro, Sushruta Mishra, Mahendra Gourisaria, Saurabh Bilgaiyan 
  Abstract: In machine learning-based SFP, software metrics and defect data are utilised as independent features, and the datasets typically have high dimensionality. To reduce dimensionality and improve model performance, an effective feature selection (FS) approach is necessary. This study proposes a new FS method called driving training-based optimisation (FSDTBO). Unlike existing FS methods, the proposed algorithm maintains a healthy balance between exploration and exploitation, thereby identifying the most relevant features for SFP. The performance of FSDTBO was compared with four widely used FS methods: FSACO, FSGA, FSDE, and FSPSO. Four classifiers NB, KNN, QDA, and DT were employed for evaluation. Experimental results demonstrate that FSDTBO achieves higher classification accuracy and selects a more optimal subset of features compared to the other methods. Moreover, statistical analysis further confirms the superiority of the proposed approach over existing FS algorithms.
  Keywords: feature selection; wrapper-based approach; software fault prediction; SFP; classifiers; driving training-based optimisation.
  DOI: 10.1504/IJCSE.2026.10078586
   
  
• SMA-CLMPNet: spatial multiscale attention enabled convolutional distributed memory network for intra-frame video forgery detection  
  by Neha Dhiman, Hakam Singh, Abhishek Abhishek 
  Abstract: Video forgery detection presents a vital threat to digital media authenticity because it allows harmful modifications through deepfakes, splicing, and frame duplication methods. Earlier techniques lack the sufficient capability to capture long-range temporal dependencies and maintaining cross-channel correlations which degrade their detection performance. Hence, this research proposes a spatial multiscale attention coupled convolutional distributed long short-term memory based modified pooling network (SMA-CLMPNet) for detecting intra-frame video forgeries. The framework process spatial, channel, and multi-scale features to capture complex forged patterns. Additionally, the distributed LSTM and modified pooling techniques in CLMPNet addresses temporal inconsistencies by selectively aggregating the most informative features while reducing the loss of critical information. The experimental results are analysed using the Face Forensics++ dataset. At the training percentage of 90%, the framework attained an accuracy of 97.92%, sensitivity of 97.34%, and specificity of 98.20%, respectively showcasing it as the prominent solution for detecting manipulations and forgeries in videos.
  Keywords: video forgery detection; digital forensics; deep learning; digital video security; feature extraction; attention mechanism.
  DOI: 10.1504/IJCSE.2026.10078605
   
  
• Graph transformer decision model for efficiently UAV-UGV joint routing problem  
  by Ke Zhang, Yuelong Su 
  Abstract: The integration of unmanned aerial vehicles with unmanned ground vehicles presents a promising solution for last-mile delivery in urban environments. However, current research often models vehicle movement as discrete nodes, rather than regards as the waypoint selection in continuous space. This paper propose a novel framework named Graph Transformer Decision Model (GTDM), to address the challenging problem of joint routing problem for multiple drones and ground vehicles. The framework employs graph transformer module for state encoding of the heterogeneous multi-agent system. Then, during the decision process, agent computes task assignment probabilities for drones through an attention mechanism and generates optimal scheduling strategies using greedy decoding. Evaluated on the real-world road network of Xiong'an New Area, results demonstrate that the proposed method successfully balances solution quality with computational efficiency compared with heuristic algorithms, such as large neighbor search, showcasing its robustness and potential for practical deployment in future smart cities.
  Keywords: unmanned aerial vehicles; vehicle routing problem; transformer; urban logistics; reinforcement learning.
  DOI: 10.1504/IJCSE.2026.10078648
   
  
• ESPS: an effective software pipelining scheme for high performance computing in vector VLIW DSP  
  by Yonghua Hu, Tang Zhuoyou, Cheng Aobo, Xiong JingCheng, Wei Liang 
  Abstract: Processors based on very long instruction word (VLIW) architecture have rich registers and computing resources. Making full use of these resources can improve code performance effectively. Software pipelining is an important optimisation technology known to improve the utilisation of computing units. However, unsuccessful instruction pipelining often appears in traditional software pipelining, so there are still many chances to further improve the arrangement of instruction sequences during the optimisation process. In this paper, to generate more efficient programs and make software pipelining optimisation more general, we present an effective software pipelining scheme (ESPS) in Vector VLIW DSP for high performance computing. The loop optimised by ESPS gets higher instruction level parallelism and more instruction scheduling space. We show that ESPS will not fail and improves the execution performance of some representative digital signal processing algorithms relative to loop unrolling optimisation by up to 5%
  Keywords: vector VLIW DSP; loop optimisation; software pipelining; high performance computing.
  DOI: 10.1504/IJCSE.2026.10078985