PeerJ Computer Science最新文献

Plagiarism detection in Arabic texts remains a significant challenge due to the complex morphological structure, rich linguistic diversity, and scarcity of high-quality labeled datasets. This study proposes a robust framework for Arabic plagiarism detection by integrating Siamese neural networks (SNN) with state-of-the-art transformer architectures, specifically AraT5 and Longformer. The system employs a hybrid workflow, combining transformer-based encoders and a classification objective to implicitly learn textual similarity. To address the inherent imbalance in Arabic plagiarism datasets, both weighted cross-entropy loss and Dice loss functions were utilized to optimize model training. Extensive experiments were conducted using the ExAraCorpusPAN2015 dataset, demonstrating the effectiveness of the proposed architecture. Results indicate that AraT5 with weighted cross-entropy loss outperformed other configurations, achieving an F1-score of 0.9058. Additionally, comparative analysis with existing methodologies highlights the superiority of our approach in handling nuanced semantic and structural variations within Arabic texts. This study underscores the importance of transformer-based architectures and class-specific loss functions in enhancing plagiarism detection accuracy in under-resourced languages like Arabic.

Terrain Aided Navigation (TAN) systems hold significant potential for delivering accurate navigation for Uncrewed Aerial Vehicles (UAVs). However, a major limitation of conventional TAN systems lies in the time-consuming correlation technique used to search the a priori map, specifically the Digital Elevation Maps (DEM). This article presents a fuzzy heuristic method for the mean absolute deviation (MAD) correlation scheme (FH-MAD), aimed at reducing the computational complexity and execution time of the TAN algorithm. The fuzzy logic system uses heading and roll angle data from onboard sensors to determine the aircraft's matching area. The output membership functions are designed based on parameters that depend on terrain features. Additionally, the proposed method incorporates an error state Kalman Filter (ESKF) as the navigation algorithm to estimate the UAV's position under various maneuvering conditions. To evaluate the effectiveness of the proposed system, tests were conducted using two distinct DEMs with varying topographical characteristics and dimensions. The results demonstrate improved position accuracy and a significant reduction in computation time compared to traditional TAN methods, making the approach suitable for real-time UAV navigation applications.

Accurate temperature and rainfall (T&R) forecasting is vital for the climate-sensitive regions of Northern India, particularly Jammu, Kashmir, and Ladakh, where volatile weather patterns significantly affect livelihoods, socio-economic development, and disaster management efforts. Despite their importance, traditional forecasting methods often fall short due to their high computational demands and inability to provide localized, real-time predictions, leaving a critical research gap in addressing these challenges. This study addresses the need for precise and efficient T&R forecasting using deep learning-based framework tailored to the unique climatic conditions of these regions. The major research focus is to develop and evaluate a model capable of capturing complex temporal dependencies in localized time-series weather data. Utilizing data from the Indian Meteorological Department (IMD) for Jammu, Srinagar, and Ladakh stations covering the period from January 1, 2000, to December 31, 2023, the proposed framework employs recurrent neural networks (RNN) and long short-term memory (LSTM) architectures, both optimized for time-series forecasting. Key findings reveal that while both RNN and LSTM models exhibit robust performance in single input single output (SISO) setups, RNN model consistently outperforms the LSTM in capturing intricate temporal relationships. The RNN model in MIMO configuration achieved significantly lower mean absolute error (MAE), root mean squared error (RMSE), and mean squared error (MSE) for Jammu, Srinagar, and Ladakh, with respective values of [0.0636, 0.1011, 0.0401] for Jammu, [0.1048, 0.1555, 0.0455] for Srinagar, and [0.0854, 0.1344, 0.0411] for Ladakh. These results underscore the RNN model's precision, making it a practical tool for real-time weather forecasting. By enhancing the accuracy of T&R predictions in regions with challenging meteorological conditions, this study contributes to improved climate adaptation strategies, disaster preparedness, and sustainable development. Its findings hold broader implications for advancing localized forecasting technologies in other regions with similar climatic complexities.

Conversational recommender systems (CRS) facilitate natural language interactions for more effective item suggestions. While these systems show promise, they face challenges in effectively utilizing and integrating informative data with conversation history through semantic fusion. In this study we present an innovative framework for extracting social information from conversational datasets by inferring ratings and constructing user-item interaction and user-user relationship graphs. We introduce a social information sensitive semantic fusion (SISSF) method that employs contrastive learning (CL) to bridge the semantic gap between generated social information and conversation history. We evaluated the framework on two public datasets (ReDial and INSPIRED) using both automatic and human evaluation metrics. Our SISSF framework demonstrated significant improvements over baseline models across all metrics. For the ReDial dataset, SISSF achieved superior performance in recommendation tasks (R@1: 0.062, R@50: 0.437) and conversational quality metrics (Distinct-2: 4.223, Distinct-3: 5.595, Distinct-4: 6.155). Human evaluation showed marked improvement in both fluency (1.81) and informativeness (1.63). We observed similar performance gains on the INSPIRED dataset, with notable improvements in recommendation accuracy (R@1: 0.046, R@10: 0.129, R@50: 0.269) and response diversity (Distinct-2: 2.061, Distinct-3: 4.293, Distinct-4: 6.242). The experimental results consistently validate the effectiveness of our approach in both recommendation and conversational tasks. These findings suggest that incorporating social context through CL can significantly improve the personalization and relevance of recommendations in conversational systems.

Detecting defects in fabrics is essential to quality control in the manufacturing process of textile productions. To increase detection efficiency, a variety of automatic fabric defect detections have been developed. However, most of these methods rely on complex model with heavy parameters, leading to high computational costs that hinder their adaptation to real-time detection environments. To overcome these obstacles, we proposed a lightweight fabric defect detection (Light-FDD), building upon the You Only Look Once v8 Nano (YOLOv8n) framework with further optimizations. Specifically, the backbone employed an improved FasterNet architecture for feature extraction. In order to capture multi-scale contextual information, we designed a parallel dilated convolution downsampling (PDCD) block to replace the conventional downsampling block in the backbone. In addition, a novel dual attention mechanism, called the global context and receptive-filed (GCRF) attention, was presented to help the model focus on key regions. Furthermore, a lightweight cross-stage partial (CSP) layer was deployed by dual convolution for feature fusion, reducing redundant parameters to further lighten the model. Results from extensive experiments on public fabric defect datasets showed that Light-FDD outperforms existing state-of-the-art lightweight models in terms of detection accuracy while requiring low computational cost. The present study suggests that the performance and effectiveness of detection models can be balanced through the adoption of reasonable strategies.

In image recognition tasks, subjects with long distances and low resolution remain a challenge, whereas gait recognition, identifying subjects by walking patterns, is considered one of the most promising biometric technologies due to its stability and efficiency. Previous gait recognition methods mostly focused on constructing a sophisticated model structure for better model performance during evaluation. Moreover, these methods are primarily based on traditional convolutional neural networks (CNNs) due to the dominance of CNNs in computer vision. However, since the alternative form of Transformer, named Vision Transformers (ViTs), has been introduced into the computer vision field, the ViTs have gained strong attention for its outstanding performance in various tasks. Thus, unlike previous methods, this project introduces two Transformer-based methods: a completely ViTs-based method GaitTriViT, and a Video Vision Transformer (Video ViT) based method GaitVViT. The GaitTriViT leverages the ViTs to gain more fine-grained spatial features, while GaitVViT enhances the capacity of temporal extraction. This work evaluates their performances and the results show the still-existing gaps and several encouraging outperforms compared with current state-of-the-art (SOTA), demonstrating the difficulties and challenges these Transformer-based methods will encounter continuously. However, the future of Vision Transformers in gait recognition is still promising.

Traditional market analysis tools primarily rely on unidimensional data, such as historical trading records and price trends. However, these data are often insufficient to reflect the actual state of the market fully. This study introduces a meta-learning-based (MLB) multimodal data fusion approach to optimize feature extraction and fusion strategies, addressing the complexity and heterogeneity inherent in international trade market data. Initially, the mel-frequency cepstral coefficients (MFCC) method is employed to transform the original audio signal into more discriminative spectral features. For image data, the convolutional block attention module (CBAM) is incorporated to capture both channel-wise and spatial attention, thereby improving the model's ability to focus on market-relevant information. In the feature fusion stage, a meta-learning bidirectional feature pyramid network (ML-BiFPN) is proposed to refine the interaction of multi-scale information via a bidirectional feature pyramid structure. An adaptive weighting mechanism is employed to adjust the feature fusion ratio dynamically. Experimental results demonstrate that the proposed multimodal data fusion model, ML-BiFPN under meta-learning, significantly outperforms existing methods in prediction performance. When tested on the publicly available Trade Map dataset, the average accuracy improves by 9.37%, and the F1-score increases by 0.0473 compare to multilayer perceptron (MLP), achieving a prediction accuracy of 94.55% and an F1-score of 0.912. Notably, under small sample conditions, the model's advantage becomes even more pronounced, with an average precision (AP) improvement of 2.79%. These findings have significant implications for international trade market forecasting and decision-making, providing enterprises with a more comprehensive understanding of market dynamics, enhancing forecasting accuracy, and supporting scientifically informed decision-making to gain a competitive edge in the marketplace.

Understanding students' emotional responses to course content and assignments is crucial for developing effective teaching strategies and improving online learning resources. To address this need, we propose a novel deep learning-based framework called BERT and BTF-IDF Integrated Framework with Deep Clustering Network (BIF-DCN), designed to accurately analyze student sentiment on educational platforms. The framework combines three key components: Bidirectional Encoder Representations from Transformers (BERT) for initial text feature extraction, Bi-level Term Frequency-Inverse Document Frequency (BTF-IDF) for enhanced feature representation, and an Improved Deep Embedded Clustering (IDEC) model for sentiment classification. BERT captures rich semantic features from student comments, which are further refined using BTF-IDF to highlight informative terms. These features are then clustered using the IDEC model to identify underlying sentiment-based topics. Experimental results show that BIF-DCN achieves higher clustering accuracy than existing IDEC-based and traditional single-model approaches on both public and self-constructed datasets. In addition to performance improvements, our method enables in-depth sentiment analysis of clustered topics, offering practical insights for optimizing teaching materials. This framework provides educators with valuable tools to better understand student needs and deliver more personalized and effective instruction, ultimately enhancing teaching quality and learner satisfaction.

The architecture of software-defined networking (SDN) involves the separation of the network control plane from the routing plane. If this initiative turns out well, it has the potential to reduce operating expenses and the duration required to provide new services in comparison to traditional networks. However, this architecture has additional security concerns, including a single point of failure that could potentially provide any user with unrestricted access to the entire network. Nevertheless, it is essential to reduce the probability of security breaches. The development of immediate intrusion detection systems (IDSs) that can quickly spot and stop malicious activities like distributed denial of service (DDoS), DoS, web-attacks, and Bot-NET is an important part of SDN architecture. Several researchers are using cutting-edge methods, such as machine learning, to investigate and elucidate the causes behind the sudden rise in attacks and abnormal behavior, but the majority of these methods are deficient in terms of flexibility and accuracy. This study proposed a lightweight method for detecting different SDN attacks from intrusion-defined networks. The lightweight long short-term memory (LSTM) network has the capability to capture temporal patterns and sequential interactions in the SDN data. It also learned important context that is efficient for feature extraction and then developed supervised random forest (SRF) for the attack prediction. The dataset consists of 207,146 rows and 84 features that were preprocessed, including separate features and target attacks. The experiments show that the proposed method achieved 99.93% accuracy for attack detection and 0.0090 loss, confirming its efficacy. We also tested the proposed method on another SDN dataset and achieved 99.43% accuracy for multi-class attack detection. Furthermore, the use of supervised random forest reduces the model's complexity, resulting in increased overall efficiency.

This article studies the multicriteria problems of scheduling a set of n products on a fabrication facility, focusing on batch availability and precedence constraints. Each product is composed of two distinct subassemblies: a common subassembly, shared across all products, and a unique subassembly unique to each product. The common subassemblies are processed together in batches, with each batch requiring an initial setup, while unique subassemblies are handled individually. The availability of a common subassembly is contingent upon the completion of its entire batch (i.e., batch availability), whereas a unique subassembly becomes available immediately after its processing. The product completion time is determined by the availability of both subassemblies. Strict (weak) precedence means that if a product precedes another, then the latter can start only after the former is completed (the latter cannot start earlier than the former). We propose O(n⁴)-time algorithms to simultaneously optimize makespan and maximum cost, as well as to lexicographically optimize two maximum costs and makespan under strict or weak precedence constraints.