Search Results (422)

High-quality code representations are fundamental to code intelligence. Achieving such representations with parameter-efficient fine-tuning (PEFT) remains a key challenge. While code pre-trained models (CodePTMs) offer a robust foundation for general-purpose embeddings, current PEFT approaches face two main obstacles when adapting them: (i) they fail to adequately capture the deep structural characteristics of programs, and (ii) they are limited by the model’s finite internal parameters, restricting their ability to overcome inherent knowledge bottlenecks. To address these challenges, we introduce a parameter-efficient code representation learning framework that combines retrieval augmentation with structure-aware priors. Our framework features three complementary, lightweight modules: first, a structure–semantic dual-channel retrieval mechanism that infuses high-quality external code knowledge as non-parametric memory to alleviate the knowledge bottleneck; second, a graph relative bias module that strengthens the attention mechanism’s capacity to model structural relationships within programs; and third, a span-discriminative contrastive objective that sharpens the distinctiveness and boundary clarity of span-level representations. Extensive experiments on three benchmarks spanning six programming languages show that our method consistently outperforms state-of-the-art parameter-efficient baselines. Notably, on structure-sensitive tasks using the PLBART backbone, RS-Rep surpasses full fine-tuning, delivering a 22.1% improvement in Exact Match for code generation and a 4.4% increase in BLEU scores for code refinement, all while utilizing only about 5% of the trainable parameters. Full article

Federated Graph Neural Networks (FedGNNs) have emerged as a promising paradigm for decentralized graph learning across distributed data silos. However, the influence of underlying communication topologies on model accuracy and efficiency remains underexplored. This study presents a topology-aware benchmarking framework for federated GNNs, systematically evaluating the impact of network structure and aggregation strategy on performance and communication overhead. The proposed framework functions as a synthetic, communication-level digital twin that emulates Federated Learning interactions and topology-dependent dynamics under controlled conditions. Four learning schemes—Centralized, Local, FedAvg, and FedAvg-Fedadam—were assessed across three representative topologies: Barabási–Albert (BA), Watts–Strogatz (WS), and Erdős–Rényi (ER). Results demonstrate that centralized training achieved the highest mean ROC-AUC (0.63), while FedAvg-Fedadam attained the best F1-score (0.038), balancing local adaptation and global convergence. Among topologies, BA and WS yielded higher average AUC values (approximately 0.57 and 0.56, respectively) than ER (approximately 0.39). Communication analysis revealed FedAvg as the most efficient strategy, requiring only approximately 3.8 × 10⁵ bytes cumulatively. These findings highlight key trade-offs between accuracy, robustness, and communication efficiency in federated graph learning and provide empirical guidance for topology-aware optimization of distributed GNNs. While the experiments rely on representative synthetic topologies, the insights offer indicative relevance and potential applicability to Internet-of-Things (IoT), vehicular, and cyber-physical networks, where communication structure and bandwidth constraints critically influence collaborative intelligence. By modeling canonical connectivity patterns and releasing our code and data, the proposed benchmarking framework offers a reproducible basis for comparing emerging federated graph architectures under constrained communication conditions. Full article

The high accessibility and interconnectedness of social media platforms have led to their increasing exploitation as tools for criminal activity. A notable example of such digital sexual offenses is the “Nth Room” case, in which sexually exploitative content and illegal recordings were unlawfully distributed on platforms such as X, Telegram, and Discord. Despite amendments to legislations, including the Sexual Violence Punishment Act and Youth Protection Act, aimed at preventing the recurrence of incidents, these crimes continue to persist. Perpetrators employ tactics such as the repeated creation and deletion of accounts, which complicate efforts to track and apprehend them. Consequently, there is an urgent need to develop advanced cyber investigation technologies capable of effectively monitoring sexual crimes posted on social media. This study aimed to propose a novel cyber investigation technology designed to trace criminal organizations by collecting tweets related to sexual crimes from X, which is the most frequently used social media platform for such content in Korea, and subsequently constructing a knowledge graph. Slang terms commonly associated with sexual crimes on X were employed as search keywords to collect relevant tweets. The knowledge graph is then generated based on three key elements extracted from the tweets: hashtags, words, and URL/invite codes. This graph serves as a tool for tracking the criminal networks involved in the distribution of sexually exploitative content and unauthorized recordings. Furthermore, to enhance tracking efficiency, an optimization model was developed to generate knowledge graphs from various analytical perspectives. In this study, to evaluate the performance of the proposed technology, a dataset of 3387 tweets was collected using an X crawler. Knowledge graphs were generated and optimized through both single and combined analyses of the three key elements, demonstrating the effectiveness of the proposed technology in tracking criminal organizations engaged in sexual crimes. Full article

This paper presents new fast algorithms for the type VII discrete cosine transform (DCT-VII) applied to input data sequences of lengths ranging from 3 to 8. Fast algorithms for small-sized trigonometric transforms enable the processing of small data blocks in image and video coding with low computational complexity. To process the information in image and video coding standards, the fast DCT-VII algorithms can be used, taking into account the relationships between the DCT-VII and the type II discrete cosine transform (DCT-II). Additionally, such algorithms can be used in other digital signal processing tasks as components for constructing algorithms for large-sized transforms, leading to reduced system complexity. Existing fast odd DCT algorithms have been designed using relationships among discrete cosine transforms (DCTs), discrete sine transforms (DSTs), and the discrete Fourier transform (DFT); among different types of DCTs and DSTs; and between the coefficients of the transform matrix. However, these algorithms require a relatively large number of multiplications and additions. The process of obtaining such algorithms is difficult to understand and implement. To overcome these shortcomings, this paper applies a structural approach to develop new fast DCT-VII algorithms. The process begins by expressing the DCT-VII as a matrix-vector multiplication, then reshaping the block structure of the DCT-VII matrix to align with matrix patterns known from the basic papers in which the structural approach was introduced. If the matrix block structure does not match any known pattern, rows and columns are reordered, and sign changes are applied as needed. If this is insufficient, the matrix is decomposed into the sum of two or more matrices, each analyzed separately and transformed similarly if required. As a result, factorizations of DCT-VII matrices for different input sequence lengths are obtained. Based on these factorizations, fast DCT-VII algorithms with reduced arithmetic complexity are constructed and presented with pseudocode. To illustrate the computational flow of the resulting algorithms and their modular design, which is suitable for VLSI implementation, data-flow graphs are provided. The new DCT-VII algorithms reduce the number of multiplications by approximately 66% compared to direct matrix-vector multiplication, although the number of additions decreases by only about 6%. Full article

Smart contracts are vulnerable to critical, design-level Business Logic Flaws (BLFs) that conventional analysis tools often fail to detect. To address this semantic gap, this study introduces a novel ontological framework that formally models the link between high-level architectural intent and low-level Sui Move code. The methodology employs a rigorous Linked Open Terms (LOT) approach to construct a comprehensive ontology, integrated with a library of secure design patterns and process-aware Object-Centric Dynamic Condition Response (OC-DCR) graphs. Qualitative validation was conducted on four canonical security patterns (Access Control, Circuit Breaker, Time Incentivization, Escapability) drawn from the official Sui Framework, confirming the framework’s representational adequacy and logical consistency. Ultimately, this work contributes the first machine-readable semantic layer for Sui Move, decoupling reasoning from raw code availability, and providing the essential semantic foundation for the future development of pattern-aware auditing tools. Full article

Code search has received significant attention in the field of computer science research. Its core objective is to retrieve the most semantically relevant code snippets by aligning the semantics of natural language queries with those of programming languages, thereby contributing to improvements in software development quality and efficiency. As the scale of public code repositories continues to expand rapidly, the ability to accurately understand and efficiently match relevant code has become a critical challenge. Furthermore, while numerous studies have demonstrated the efficacy of deep learning in code-related tasks, the mapping and semantic correlations are often inadequately addressed, leading to the disruption of structural integrity and insufficient representational capacity during semantic matching. To overcome these limitations, we propose the Functional Program Graph for Code Search (called FPGraphCS), a novel code search method that leverages the construction of functional program graphs and an early fusion strategy. By incorporating abstract syntax tree (AST), data dependency graph (DDG), and control flow graph (CFG), the method constructs a comprehensive multigraph representation, enriched with contextual information. Additionally, we propose an improved metapath aggregation graph neural network (IMAGNN) model for the extraction of code features with complex semantic correlations from heterogeneous graphs. Through the use of metapath-associated subgraphs and dynamic metapath selection via a graph attention mechanism, FPGraphCS significantly enhances its search capability. The experimental results demonstrate that FPGraphCS outperforms existing baseline methods, achieving an MRR of 0.65 and ACC@10 of 0.842, showing a significant improvement over previous approaches. Full article

This article presents a novel methodology for assessing the visibility and reachability of cultural heritage objects within urban structures, tested through a pilot study in Kaunas New Town (Naujamiestis), Lithuania. While heritage protection policies usually emphasize architectural composition, details, and external visual protection zones, interior urban views and functional spatial dynamics remain underexplored. Building upon Space Syntax theory and John Peponis’s concepts of distributive and correlative situational codes, this study integrates detailed visibility analysis with graph-based accessibility modeling. Visibility was quantified through a raster-based viewshed analysis of building footprints and street-based observation points, producing a normalized visibility index. Reachability was examined using a new graph indicator based on the ratio of reachable polygon area to perimeter (A²/P), further weighted by the area of adjacent buildings to reflect the potential for urban activity. Validation against independent datasets (population, companies, and points of interest) confirmed the superior explanatory power of the proposed indicator over traditional centralities. By combining visibility and reachability in a bivariate matrix, the model provides insights into heritage objects’ dual roles as landmarks, everyday hubs, or hidden sites, and offers predictive capacity for evaluating urban transformations and planning interventions. Full article

Advancements in natural language processing (NLP), particularly Large Language Models (LLMs), have greatly improved how we access knowledge. However, in critical domains like biomedicine, challenges like hallucinations—where language models generate information not grounded in data—can lead to dangerous misinformation. This paper presents a hybrid approach that combines LLMs with Knowledge Graphs (KGs) to improve the accuracy and reliability of question-answering systems in the biomedical field. Our method, implemented using the LangChain framework, includes a query-checking algorithm that checks and, where possible, corrects LLM-generated Cypher queries, which are then executed on the Knowledge Graph, grounding answers in the KG and reducing hallucinations in the evaluated cases. We evaluated several LLMs, including several GPT models and Llama 3.3:70b, on a custom benchmark dataset of 50 biomedical questions. GPT-4 Turbo achieved 90% query accuracy, outperforming most other models. We also evaluated prompt engineering, but found little statistically significant improvement compared to the standard prompt, except for Llama 3:70b, which improved with few-shot prompting. To enhance usability, we developed a web-based interface that allows users to input natural language queries, view generated and corrected Cypher queries, and inspect results for accuracy. This framework improves reliability and accessibility by accepting natural language questions and returning verifiable answers directly from the knowledge graph, enabling inspection and reproducibility. The source code for generating the results of this paper and for the user-interface can be found in our Git repository: https://git.zib.de/lpusch/cyphergenkg-gui, accessed on 1 November 2025. Full article

Plant microRNAs (miRNAs) are endogenous non-coding RNAs (~20–24 nucleotides) that regulate gene expression post-transcriptionally, playing critical roles in plant growth, development, and stress responses. This review systematically examines AI applications in plant miRNA research, tracing evolution from traditional machine learning to deep learning architectures. Plant miRNAs exhibit distinctive features necessitating plant-specific computational approaches: nuclear-localized biogenesis, high target complementarity (>80%), and coding region targeting. These characteristics enable more accurate computational prediction and experimental validation than animal systems. Methodological advances have improved prediction accuracy from ~90% (early SVMs) to >99% (recent deep learning), though metrics reflect different evaluation contexts. We analyze applications across miRNA identification, target prediction with degradome validation, miRNA–lncRNA interactions, and ceRNA networks. Critical assessment reveals that degradome data capture mixed RNA fragments from multiple sources beyond miRNA cleavage, requiring stringent multi-evidence validation. Similarly, fundamental ambiguities in lncRNA definition compound prediction uncertainties. Major challenges include severe data imbalance (positive to negative ratios of 1:100 to 1:10,000), limited cross-species generalization, insufficient model interpretability, and experimental validation bottlenecks. Approximately 75% of plant miRNA families in miRBase v20 lack convincing evidence, underscoring the need for rigorous annotation standards. Future directions encompass multimodal deep learning, explainable AI, spatiotemporal graph neural networks, and ultimately AI-driven de novo miRNA design, though the latter requires substantial advances in both computation and high-throughput validation. This synthesis demonstrates that AI has become indispensable for plant miRNA research, providing essential support for crop improvement while acknowledging persistent challenges demanding continued innovation. Full article

Graph generalization is a powerful concept with a wide range of potential applications, while established algorithms exist for generalizing simple graphs, practical approaches for more complex graphs remain elusive. We introduce a novel formal model and algorithm (GGA) that generalizes labeled directed graphs without assuming label identity. We evaluate GGA by focusing on its information preservation relative to its input graphs, its scalability in execution, and its utility for three applications: abstract syntax trees, class graphs, and call graphs. Our findings reveal the superiority of GGA over alternative tools. GGA outperforms ASGard by an average of 5–18% on metrics related to information preservation; GGA matches 100% with diffsitter, indicating the correctness of the output. For class graphs, GGA achieves 77.1% in precision at 5, while for call graphs, it exhibits 60% in precision at 5 for a specific application problem. We also test performance for the first two applications: GGA’s execution time scales linearly with respect to the product of vertex count and edge count. Our research demonstrates the ability of GGA to preserve information in diverse applications while performing efficiently, signaling its potential to advance the field. Full article

The application of graph convolutional neural networks for traffic prediction is a standard procedure; however, this approach is rarely used under the assumption that the exact city plan is unknown and the prediction area is a city-sized region. This paper fills this gap by proposing and evaluating the Sample and Aggregate-Voronoi method (SAGE-Voronoi), which utilizes the novel concept of Voronoi Neighborhood Weighted Graph-based convolutional networks to predict car traffic in cities. It demonstrates the usefulness of this method for short-term predictions using real sensor data from the moderate-sized town of Darmstadt. The results obtained are compared with those of other neural network algorithms, namely pure Long Short-Term Memory, SAGE, Diffusion Convolutional Gated Recurrent Unit (DCGRU), and Spatio-Temporal Graph Convolutional Neural Network (STGCN). SAGE-Voronoi obtained significantly better results than the state-of-the-art approaches. The SAGE-Voronoi graph neural network enables the reliable prediction of varying car traffic among network nodes. The proposed approach is not limited to spatiotemporal traffic data and can be utilized in other similar domains. The source code and dataset used in our experiments are available for download, enabling full reproducibility of the results. Full article

Large Language Models of the Transformer architecture display great promise in automated code error detection based on their strength in processing sequential data. Nevertheless, their efficacy could be further improved by addressing the inherent weakness in handling structural code dependencies. In response to this, we introduce a novel model that integrates the semantic comprehension power of RoBERTa with the structural learning strength of Graph Neural Networks. This model aims to detect the most common categories of programming faults in the form of runtime errors, index errors, and import/module errors. Experimental evaluation has demonstrated that the hybrid model, utilizing a proper fusion technique, outperforms other models in terms of accuracy and robustness. The introduced mechanism leads to numerical benefits, improving test accuracy by 1.75% over competitive baseline. Full article

As software systems grow increasingly complex and interconnected, detecting vulnerabilities in source code has become a critical and challenging task. Traditional static analysis methods often fall short in capturing deep, context-dependent vulnerabilities and adapting to rapidly evolving threat landscapes. Recent efforts have explored knowledge graphs and transformer-based models to enhance semantic understanding; however, these solutions frequently rely on static knowledge bases, exhibit high computational overhead, and lack adaptability to emerging threats. To address these limitations, we propose DynaKG-NER++, a novel and lightweight framework for context-aware vulnerability detection in source code. Our approach integrates lexical, syntactic, and semantic features using a transformer-based token encoder, dynamic knowledge graph embeddings, and a Graph Attention Network (GAT). We further introduce contrastive learning on vulnerability–patch pairs to improve discriminative capacity and design an attention-based fusion module to combine token and entity representations adaptively. A key innovation of our method is the dynamic construction and continual update of the knowledge graph, allowing the model to incorporate newly published CVEs and evolving relationships without retraining. We evaluate DynaKG-NER++ on five benchmark datasets, demonstrating superior performance across span-level F1 (89.3%), token-level accuracy (93.2%), and AUC-ROC (0.936), while achieving the lowest false positive rate (5.1%) among state-of-the-art baselines. Sta tistical significance tests confirm that these improvements are robust and meaningful. Overall, DynaKG-NER++ establishes a new standard in vulnerability detection, balancing accuracy, adaptability, and efficiency, making it highly suitable for deployment in real-world static analysis pipelines and resource-constrained environments. Full article

Precise, seamless, and high-rate navigation remains a major challenge, particularly when relying on low-cost sensors. With the decreasing cost of cameras, Inertial Measurement Units (IMUs), and Global Navigation Satellite System (GNSS) receivers, tightly coupled fusion frameworks, such as GVINS, have gained considerable attention. GVINS is an optimization-based factor-graph framework that integrates visual and inertial measurements with single-frequency GNSS-code pseudorange observations to provide robust and drift-free navigation. This study aimed to evaluate the potential of applying GVINS to low-cost, low-power, and single-frequency GNSS receivers, particularly those embedded in smartphones, by integrating 1 Hz GNSS measurements collected in three challenging urban scenarios into the GVINS framework to produce seamless 10 Hz positioning estimates. The experiments were conducted using an Xsens MTi-1 IMU and global-shutter (GS) cameras, as well as a Samsung A51 smartphone and a u-blox ZED-F9P GNSS receiver. GVINS was modified to process 1 Hz GNSS measurements. Differential corrections from a nearby GNSS reference station were also incorporated to assess their impact on optimization-based filters, such as GVINS. The performance of GVINS and Differential GVINS (D-GVINS) solutions using smartphone measurements was compared against standard point positioning (SPP) and differential GPS (DGPS) results obtained from the same smartphone GNSS receiver, as well as the GVINS solution derived from u-blox ZED-F9P measurements sampled at 1 Hz. Experimental results show that GVINS effectively operates with smartphone GNSS measurements, reducing 3D RMS errors by 80.4%, 64.9%, and 83.8% for the sports field, campus-walking, and campus-driving datasets, respectively, when differential corrections are applied relative to the SPP solution. These results highlight the potential of smartphone GNSS receivers within the GVINS framework: Even though they observe fewer constellations, lower signal quality, and a lower number of satellites, they can still achieve a performance comparable to that of a relatively higher-end dual-frequency GNSS receiver, the u-blox ZED-F9P. Further studies will focus on adapting the GVINS algorithm to run directly on smartphones to utilize all the available measurements, including the camera, IMU, barometer, magnetometer, and additional ranging sensors. Full article

This article presents a part of the industrial metaverse for electric drive system diagnostics. The advantages of using a low-code/no-code platform for electric drive systems diagnostics are demonstrated. Five diagnostic scenarios were developed, programmed, and implemented. The article demonstrates the implementation and use of the platform’s main functional blocks: a visualization block (which displays the state of electric machines in any user-friendly form—graphs, Park’s vector diagrams, or diagnostic curves); a digital twin block (which simulates various engine states); a digital twin block with an engine defect (which simulates faulty engine states); and an artificial intelligence block (which trains classification model to predict various engine states). Experiments on training the artificial intelligence block using a misalignment defect dataset are presented. The dataset was divided into six classes: engine operation with/without a defect under no load, engine operation with/without a defect under a 50% load, and engine operation with/without a defect under a 100% load. The workflow for training and using the model, the basic training approaches, and the distinguishability of the presented classes are demonstrated. The model training results are shown. The article presents a methodology for extensive testing of program functionality. The obtained results demonstrate the feasibility of implementing a low-code/no-code platform and the feasibility of solving the assigned tasks with its help, as well as the simplification and reduction in engineering solution development time. Full article