Vol. 45 · No. 2

Software Re-Engineering (SRE) and Sustainable Software Engineering (SSE) are distinct yet interrelated paradigms in modern software development. SRE focuses on restructuring and optimizing existing software systems to improve functionality, maintainability, and adaptability, while SSE emphasizes incorporating environmentally responsible practices throughout the software life cycle. This paper presents a comparative analysis of these two approaches, highlighting their foundational principles, methodologies, challenges, and benefits. It underscores the necessity of embedding green and resource-efficient practices into re-engineering activities to achieve both operational excellence and environmental sustainability. By integrating SRE with SSE, organizations can enhance software quality, reduce ecological impact, and ensure long-term viability within the evolving software ecosystem. The study further advocates for energy-efficient design, user engagement, and renewable energy integration in software infrastructure. The findings aim to guide practitioners and researchers toward developing software systems that are not only optimized for performance but also aligned with sustainability goals, thus bridging the gap between technological advancement and environmental stewardship.

With the continuous development of smart mobile devices, user check-in behaviors data has experienced explosive growth. To solve the problem that existing point-of-interest (POI) recommendation models fail to fully consider the complex and dynamic changes in user preference behaviors, a POI recommendation model based on a Bi-directional Long Short-Term Memory recurrent neural network (Bi-LSTM) and an attention mechanism is proposed. By integrating the time sequence and spatial location information of user check-in behaviors into Bi-LSTM, the user's spatio-temporal feature sequential dependence is established. The attention mechanism is utilized to model historical check-in locations, derive a weight sequence, and recommend the next POI to the user based on the probability distribution of predicted POIs. Experimental results on two public datasets demonstrate that the model outperforms previous advanced POI recommendation models in multiple evaluation metrics and has better recommendation performance.

The traditional collaborative filtering recommendation model usually cannot capture the correlation between user items accurately when faced with the problem of data sparsity, which leads to a poor recommendation effect. To solve these problems, this paper proposed a recommendation model integrating semantic features and knowledge features. By introducing a knowledge graph as auxiliary information, the problem of data sparseness was alleviated, and the granularity of user interest description was refined based on the rich knowledge information. Different from other recommendation models based on knowledge graphs, the model in this paper mined implicit feedback information between user projects based on semantic features and knowledge features. The Word2vec model was used to capture the semantic features of project entities, and knowledge features were learned according to the idea of preference diffusion to obtain the user preference diffusion set integrating features of different levels and forming the user preference feature sequence from the inner layer to the outer layer. The sequence was adopted as the input of the gated cycle unit, and the attention mechanism was integrated to capture the deep feature information between the user and the item, so as to enrich the user representation to improve the recommendation effect. The comparison experiment of the CTR prediction scenario and the Top-K recommendaton scenario, and the sparseness verification experiment were all conducted on the MovieLens dataset. The results showed that, compared with BPR-MF, DKN and RippleNet baseline models, this model had better performance in both scenarios and could effectively alleviate the impact of data sparseness.

Prioritising user stories is a fundamental challenge in requirements engineering. It directly impacts project outcomes and is a key factor in agile success. Popular techniques, such as MoSCoW, rely on subjective and static criteria and do not expose dependencies, which can lead to suboptimal sprint and release plans. This article presents a scenario-based prioritisation technique grounded in identified usage scenarios. The technique considers contextual factors, dependencies, and the business impact of delivered user stories. We outline the concept, illustrate its mechanics on a case example, and compare results against classic prioritisation. Preliminary findings indicate that scenario-based prioritisation can serve as a valuable complementary tool for Product Owners when selecting requirements critical to end users. We observed limited overlap between scenario-based priorities and MoSCoW must-haves, with only modest alignment to value-based rankings. A brief survey also indicated a preference for a scenario-oriented MVP over a set of independent high-value features. We discuss implications for planning coherent increments and conclude with directions for further work, including larger-scale evaluations, real-world experiments, and dedicated tool support.

The database management system based on Machine Learning (ML) algorithm can achieve effective classification by acquiring and analyzing massive data. Therefore, this paper analyzed and studied the Big Data (BD) information management and database construction based on ML algorithm. Artificial intelligence provided a feasibility analysis for this article. The research and analysis results showed that the BD information management efficiency based on ML algorithm was 10.17

Phishing remains a cybersecurity risk by exploiting human nature and technology loopholes to steal sensitive data. In this research, with secondary qualitative analysis including academic databases and content analysis, machine-learning-based phishing models with multidimensional feature analysis are critically analyzed. Computational inefficiency, real-time detection capability, and ethics concerns like model explainability and privacy are some of the challenges. Model reliability is the primary issue of feature ablation analysis of features like URL structure, domain registration, and webpage behavior as well as comparative performance evaluation of classifiers like Naive Bayes, SVM, Random Forest, Decision Trees, and Gradient Boosting. Scalable, adaptive, and ethical phishing is proposed using NLP and blockchain-based hybrid solutions for long-term digital pathology systems.

Deep Learning-based electroencephalogram (EEG) decoding in brain-computer interface (BCI) faces many challenges: First, EEG has low spatial resolution, which makes it difficult to accurately identify specific brain regions; Second, deep learning requires large amounts of data, while EEG datasets are typically small. Merging small datasets raises privacy concerns because EEG signals contain a wealth of personal information. Finally, while federated learning (FL) offers a potential solution for addressing the issues of limited data and privacy, it faces the challenge of non-independent and identically distributed (non-IID) data, as the local data on the client is often derived from datasets of different tasks or subjects. To solve these problems, this paper proposes a novel Federated Graph Learning (FGL) framework based on Electroencephalogram and Electromyography (EEG-EMG), named FSDFGL. During the construction of the graph data, EMG data is incorporated into EEG data through SPMI, as EMG signals provide distinct features of muscle activity. In the local Graph Neural Network (GNN) process, the structural information is decoupled from the node features. Then, the structure encoder is used to capture the structural information and share it globally during FL. This allows FSDFGL to capture more commonly shared structural information while avoiding degradation of local model performance due to sharing heterogeneous features. The framework was evaluated on the TAN dataset, and experimental results demonstrate the superiority of FSDFGL in identifying complex motion intentions.

Computerized segmentation and classification of brain tumors are essential for diagnosis, management, and prediction of the prognosis in neuro-oncology. Standard methodologies for the diagnosis of brain tumors are subjective, based on human experience and qualitative analysis of images, which makes them unsuitable for handling all the variability in tumor types. To address these challenges, this research presents an innovative approach of combining DCNs and PECT images to accurately and quickly segment and classify multi-class brain tumor. The framework uses a DenseNet architecture tailored for high-resolution functional imaging in PECT to yield rich feature representations for tumor identification as well as grading and staging for several types of tumors. The computer algorithm was trained on PECT scans of 250 patients with gliomas, meningiomas and metastatic brain tumors as a training set and test set. Proper segmentation of tumor boundaries, type of tumor, and other important structures were carried out using the multi-class segmentation maps. By implication, the experimental results show that the proposed model has a classification accuracy of 94

With the continuous advancement of medical technology, medical image segmentation plays a pivotal role in various aspects of disease diagnosis, treatment planning, and therapeutic evaluation. In recent years, methods based on U-shaped Convolutional Neural Networks (CNNs), U-shaped Transformer architectures, and U-shaped Mamba models have achieved significant progress in this domain, greatly promoting research and application development. Although CNNs excel at local feature extraction, they are inherently limited in modeling global dependencies due to the locality of convolution operations. Transformers address this limitation through self-attention mechanisms that capture long-range dependencies effectively, but their quadratic computational complexity with respect to sequence length severely impacts efficiency. While the Mamba architecture demonstrates promising performance in long-sequence modeling, its generalization to other tasks remains limited. To address these challenges, this paper proposes RCLD-UNet, an efficient asymmetric U-shaped encoder-decoder network designed for 2D medical image segmentation. Built upon the U-Net backbone, RCLD-UNet incorporates the Recurrent Criss-Cross Attention (RCCA) module in the encoder, which leverages sparse connections and recurrent operations to efficiently capture global contextual information, thereby overcoming the limitations of standard convolutions in modeling global dependencies. In the decoder, a Linear Deformable Convolution (LDConv) module is introduced to enhance feature extraction through irregular convolutions, while ensuring linear parameter growth to balance performance and computational cost. Extensive experimental results on three public datasets indicate that RCLD-Unet achieves superior performance compared to existing methods. Specifically, the proposed model achieves average Dice Similarity Coefficients (DSC) of 84.14

Vision-and-language navigation faces critical challenges, including insufficient long-horizon reasoning, misalignment between instruction segments and visual observations during navigation, and hallucination interference in large model reasoning. These limitations hinder the ability of large models to accurately interpret complex scenes through textual summaries alone, particularly in understanding spatial distributions of environmental elements and dynamically adapting to the demands of target instructions. To address these challenges, we propose a novel historical-topological vision-and-language navigation framework based on multi-curvature spatiotemporal Chain-of-Thought reasoning. Leveraging an encoder-decoder architecture, our model encodes historical information from input features, effectively resolving the lack of global topological guidance in long-horizon reasoning tasks. To mitigate hallucination in large model reasoning, we exploit the inherent data characteristics of vision-and-language navigation tasks. By leveraging multi-curvature manifolds, we identify salient spatiotemporal references in visual observation sequences, enabling the model to ground its reasoning in these references for accurate situational understanding and future prediction. To enhance spatial structure perception, we introduce a graph self-attention mechanism that jointly models node distances and visual similarity in historical topological graph embeddings, effectively capturing spatial relationships between nodes. This work introduces novel insights for large model-based vision-and-language navigation in indoor scenes, advancing their applicability in complex environments.