The rapid growth of blockchain-based applications (BoS) has transformed multiple sectors but also brought significant challenges in software testing, particularly for decentralized applications (DApps) and smart contracts. Current development tools primarily support unit testing and do not address the full range of testing needs for blockchain software. Given the complexity of DApps and the immutability of smart contracts, more advanced methods are required to ensure security, performance, and functional accuracy. This study reviews the current state of blockchain software testing, identifying major gaps and limitations in conventional testing frameworks. To address these challenges, we propose an innovative software testing framework that integrates machine learning to offer real-time, customized testing recommendations for blockchain applications. By utilizing key blockchain features, including distributed ledgers, cryptographic hashing, and decentralized consensus, our model enhances testing accuracy by identifying potential vulnerabilities, performance limitations, and functional discrepancies, reducing the risk of undetected defects in the immutable blockchain environment. Experimental assessments show substantial improvements in testing coverage when compared to established tools such as Truffle and Remix, particularly in the validation of smart contracts and identification of security vulnerabilities. Our framework accelerates the testing process and improves the reliability of blockchain applications by providing developers with comprehensive tools to address the unique challenges of decentralized systems. As blockchain continues to advance in sectors like finance, healthcare, and supply chain management, this study highlights the urgent need for sophisticated testing methods and establishes a foundation for future advancements. By combining machine learning with blockchain testing, we introduce a scalable and adaptable approach that can progress alongside developments in blockchain technology. The conclusion explores broader implications and suggests further improvements, including integration into active deployment pipelines and real-time testing in operational settings. This framework marks a significant advancement in ensuring the dependability, security, and scalability of decentralized blockchain applications, supporting the sustainable growth of these systems in the digital ecosystem.
Cloud-based enterprise resource planning (ERP) systems are replacing traditional on-premise ERP for many companies, in pursuit of cloud benefits such as usage-based licensing, modernized technology and lower costs. While ERP vendors shape their strategy in a software-as-a-service (SaaS) model and promise their customers the aforementioned benefits if they migrate to a subscription-based cloud ERP, we analyse the reality behind these claims and how it affects ERP life cycle costs. This paper focuses on evaluating the total cost of ownership (TCO) between similar IaaS and SaaS systems to provide insight on their key differences. We conduct a multiple case study on four large enterprises that have moved to a SaaS model cloud ERP by analysing archival documents that reflect the actual cost differences between the old and new systems, and provide theoretical implications based on the results. Our findings indicate that the SaaS delivery model can provide cost advantages in specific cases, but IaaS could provide more value for large companies through cost efficiency and more control over the platform vendor. The findings help managers to align their cloud strategy and better understand the possible benefits and pitfalls of subscription-based cloud ERP solutions.
Web usability has become a critical success factor across many organizations in the era of web-based apps. Government websites play a crucial role in distributing information and providing services. However, many websites need more intricate navigation systems and subpar user experiences. BUTTMKP, a governmental agency entrusted with overseeing agricultural quarantine, has recently undertaken a website makeover initiative to address customer feedback and issues. This research's primary objective is to thoroughly evaluate the usability of both the old and revised BUTTMKP web interfaces. The goal is to identify areas that may be enhanced and to recognize the strengths of each design, all from a user-centric standpoint. The research utilizes the System Usability Scale (SUS) and qualitative data-gathering techniques to acquire user input and examine their experiences. The average SUS scores for the old and revised designs stood at 58.5 and 57.8, respectively, and serve as a foundation for our ongoing commitment to improving our model. We recognize that user-friendliness is a crucial aspect of any interface, and these scores encourage us to refine our design further. In the future, we will continue to enhance this model to ensure that it meets our users' evolving needs and expectations.
The application of artificial intelligence (AI) has seen a marked increase in interest, particularly within the Banking, Financial Services, and Insurance (BFSI) sector. Despite this growing interest, many stakeholders remain resistant to adopting AI due to the opaque nature of its outputs, often referred to as a "black box." Explainable Artificial Intelligence (XAI) emerges as a crucial solution, aiming to enhance the transparency and interpretability of AI outputs. Therefore, ensuring the explainability of machine learning outputs is essential in the development of machine learning models. Nevertheless, significant challenges remain in the development of machine learning software applications (MLSA), primarily due to the lack of integration between the machine learning development workflow and the Software Development Life Cycle (SDLC). This disconnect leads to inefficiencies in the MLSA development process and can ultimately result in project failures. This study proposes a software process framework that integrates the machine learning development workflow with the SDLC while ensuring that model outputs are clearly explained in alignment with the needs of MLSA stakeholders. Such integration is necessary to improve MLSA development's overall effectiveness and ensure that the machine learning models' outputs meet the specific requirements and expectations of the stakeholders involved.
The rapidly evolving video game industry faces the challenge of managing increasingly complex development projects, often involving interdisciplinary teams. These teams frequently operate remotely, at least partially, and typically rely on the Scrum project management framework, which is an adaptive approach within the Agile methodology. However, Scrum was originally designed for small, co-located teams with direct, face-to-face communication, making it less effective in the context of virtual teams. The transition to remote work has highlighted several challenges, particularly regarding communication and trust among team members, which are not adequately addressed by the traditional Scrum tools. To address these issues, this paper introduces a novel software tool that integrates graph theory to enhance the Scrum project management process. Recognizing that chat platforms, such as Microsoft Teams, are the primary mode of communication in remote work environments, our tool leverages data from these platforms. It performs both quantitative and qualitative analyses of various graph-based metrics to assess the health of team communication and provide actionable feedback to managers. This tool, developed in Python, has been tested using synthetic communication scenarios generated by Chat GPT.
Career planning is crucial for university students, not only aiding them in securing employment but also in achieving their career aspirations and life goals. However, traditional career guidance methods often fall short in providing personalized and dynamic support. This study aims to offer tailored career recommendations by leveraging a multi-modal deep learning model that integrates text, image, and behavioral data. Techniques such as Natural Language Processing (NLP), Convolutional Neural Networks (CNNs), and Long Short-Term Memory (LSTM) networks are employed to analyze diverse data sources. Comparative experiments with traditional methods demonstrate that the proposed model achieves higher accuracy and user satisfaction. The study highlights the potential of advanced data analysis techniques in enhancing career guidance services, ultimately aiding students in making well-informed career choices.
The research on Animal Species Classification using Convolutional Neural Networks (CNNs) addressed challenges such as manual identification methods, limited dataset sizes, image occlusions, and variations in lighting conditions. The study aimed to investigate the CNN algorithm for Animal Species Classification, develop a CNN based classification system, and evaluate the system's accuracy in classifying animal species. The methodology included collecting an image dataset, preprocessing the images through standardization and data augmentation, and partitioning the data into training, testing, and validation sets. A pre-trained model, MobileNet, served as a base architecture for feature extraction. The CNN architecture was meticulously designed, considering layers, activation functions, and model complexity. Additionally, a user-friendly graphical interface was created to enhance user interaction. The final results indicate a training accuracy of approximately 0.9068 with a training loss of 0.2818. The validation accuracy and loss were 0.9350 and 0.2043, respectively, while the testing accuracy was 0.91. The implications of this research are significant, potentially impacting biodiversity conservation, ecological studies, educational tools, and wildlife monitoring. Future improvements could involve expanding the dataset, fine-tuning the model to increase accuracy, and exploring applications in protecting endangered species and monitoring habitats.
Integrating AI with IoT, specifically leveraging the Edge platform enhanced with cloud computing capabilities, enables real-time data analytics and dynamic health management directly at the network edge. This report presents an innovative AIoT framework for tackling obesity in New South Wales. This hybrid system allows for scalable and efficient processing of large data sets, ensuring timely interventions tailored to individual health profiles. By continuously analyzing data collected from IoT devices, our approach provides predictive insights and personalized treatment plans, which are crucial for managing complex health conditions like obesity. The proposed solution addresses significant challenges such as data security, privacy, and the technical complexities associated with AI and IoT integration, offering a scalable model that can be adapted to other health conditions. Pilot studies will validate the framework's effectiveness, with iterative refinements enhancing reliability and user engagement. This research marks a significant step towards transforming traditional health management into a more proactive, personalized healthcare system using AIoT technologies.
Alzheimer's disease (AD) is a progressive brain disorder that leads to a decline in cognitive and functional abilities. It is one of the most common causes of dementia, with far-reaching consequences not only for patients and caregivers but also globally. Although memory loss is often regarded as the hallmark of Alzheimer's disease, language impairment can also manifest in the early stages. Early diagnosis is crucial, as therapeutics can delay the progression of the disease and provide those diagnosed with valuable time. In this work, Natural Language Processing (NLP) approaches were utilized to classify dementia patients’ spontaneous speech from the DementiaBank dataset to predict dementia. Additionally, Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) models were employed to differentiate between dementia patients and control subjects. It was found that using pre-trained GloVe embeddings significantly improved accuracy compared to random embeddings when applying the mentioned models. The transfer learning technique for text classification yielded more promising results than training the entire dataset on a large number of neural network model parameters. Performance indicators such as precision, recall, F1 score, and specificity were evaluated. Furthermore, the incorporation of the attention mechanism into the model, along with hyperparameter optimization of the CNN-LSTM model, resulted in excellent accuracy.
Camouflaged Object Detection (COD) aims to segment objects that are visually integrated into their surroundings. Despite remarkable progress, existing methods still struggle with the dual challenges of omitting detail discrimination and encountering feature redundancy, thereby failing to achieve optimal performance. Addressing these problems, we propose a novel network architecture: Wavelet Boost Identification-based Multi-level Refinement Network (WBRNet) for COD. WBRNet employs Wavelet Discrimination Boost to enhance feature recognition capabilities and effectively suppress noise. It applies distinct feature extraction strategies to high-frequency and low-frequency regions, focusing particularly on detail-rich high-frequency areas to overcome detail discrimination omission. WBRNet then integrates the Mamba framework with asymmetric convolutions, providing a more extensive receptive field. Finally, utilizing our proposed Feature Reversal Decoder, WBRNet precisely directs attention to specific image patches and generates counter-masks, enhancing the accuracy of feature extraction. Comparative results across multiple benchmark datasets demonstrate that the WBRNet outperforms 13 state-of-the-art (SOTA) methods with remarkable outcomes.
This paper proposes and implements a cloud-native heterogeneous resource scheduling platform based on Kubernetes and Docker, aimed at optimizing the scheduling and utilization of GPU resources in cloud environments. The platform integrates Kubernetes' container orchestration capabilities, NVIDIA GPU device plugins, and a custom scheduling algorithm to achieve dynamic management of GPU resources, effectively balancing the load of GPU-intensive tasks. Furthermore, the platform supports hybrid scheduling of GPUs and CPUs, optimizing task allocation among heterogeneous resources and enhancing overall computational efficiency. The system incorporates Prometheus and Loki for real-time monitoring and log management, assisting administrators in maintaining resource usage, detecting bottlenecks, and ensuring system stability. The platform leverages Kubernetes to provide self-healing capabilities, ensuring high availability in the event of node or container failures. Through extensive testing and validation, the proposed platform demonstrates outstanding performance in terms of functionality, efficiency, and fault tolerance, making it suitable for application scenarios involving GPU-intensive tasks such as deep learning, scientific computing, and image rendering.
Session-based recommendation (SBR) aims at recommending the next possible item for anonymous users based on short interacted sequences. Existing work exploits the advantages of graph neural networks in modeling higher-order relationships of neighboring items, but they tend to neglect the consideration of potentially noisy neighbors, which can cause recommendation bottlenecks. To solve this problem, we propose the Denoised Edge Aware Network (D-EAN) for SBR, which eliminates the impacts of noise items to achieve more reasonable representation learning outcomes for each item. In D-EAN, We model the interaction sessions as graphical structures and assign sparse considerations to the relationships between neighboring items to minimize the impact of disjoint neighbors on the learning of the central node. Besides, we refine the relation of item transition in the session as four different types in order to extract the finer-grained joint item relationships. Experiments conducted on three real-world datasets demonstrate that D-EAN outperforms the state-of-the-art baselines.
Network segmentation is an important cybersecurity strategy that aims to reduce an attacker's ability to move laterally and increase security within a network. This study evaluated the effectiveness of network segmentation by simulating attacks on two networks using Infection Monkey. The objective of this study was to examine the impact of network segmentation on the detection and spread of threats across several virtual machines. Two network environments were created, the first unsegmented network (Network A) and a segmented network (Network B). Multiple simulated attacks were conducted on both networks. The attacks confirmed that the segmented network achieved better protection because the simulated malware did not detect any of the virtual machines in within that network. However, all virtual machines in the unsegmented network were compromised. Network segmentation proves to be an effective method for enhancing security by preventing the detection and spreading of threats across virtual machines. This research showcases the importance of implementing segmentation in network security strategies. Future research should assess the impact of segmentation options and differentiate between them to defend against different types of threats.
In the medicine field, abbreviations are widely used for efficient text recording and simplification. However, one abbreviation can have many different meanings depending on its context and usage while one word or phrase, conversely, can be abbreviated in several various ways. Such situations cause inconsistencies and confusion for both human and machine-based processing when medical data is shared. It is even more challenging with inconsistencies and non-standardization of abbreviations in Vietnamese medical texts, especially clinical texts in hospitals, when language characteristics are unique and data and system resources are not available. Therefore, constructing an abbreviation dictionary system to support Vietnamese clinical text processing is crucial and focused in this paper. In particular, we propose the first abbreviation dictionary system that includes a dictionary of abbreviations with meanings and related information from Vietnamese medical literature. This system is constructed in a flexible evolvable manner. It supports end-users via a web-based interface and connects all other computer-based systems via Application Programming Interfaces. In addition, we develop an extraction method to automatically extract abbreviation-long form pairs from Vietnamese medical literature. The performance of our extraction method is promising with an accuracy of 98.68% as evaluated by medical experts. As a result, our abbreviation dictionary system is helpful to aid in understanding clinical texts and ensure consistent use of abbreviations. Furthermore, it can serve other applications like translation, concept extraction, medical data analysis, and decision support in medicine.
The teaching profession in the Philippines is highly esteemed, with a growing demand for college instructors due to local governments’ efforts to provide affordable education. Public colleges require instructors to hold a Master's degree, but the high demand often leads to hiring under-qualified teachers, including fresh graduates. Instructors without a Master's degree must enroll in a graduate program and complete it within three years while maintaining high performance and attendance standards. Despite these challenges, the quality of education must be prioritized, necessitating regular faculty assessments each term to ensure students receive high-quality education. To standardize faculty retention, this research develops a faculty retention ranking model using fuzzy logic. Fuzzy expert systems, which handle decision-making through fuzzy reasoning, are employed due to their effectiveness in managing uncertainty. This approach demonstrates high recognition rates and reduced computational complexity, making it suitable for developing an efficient faculty retention ranking system.
In recent years, sequential recommendation systems have focused on modeling users' historical interactions to capture their dynamic preferences, thereby achieving more accurate and personalized recommendations. Current approaches typically utilize user and item IDs along with textual features for sequence modeling, employing self-attention mechanisms or Fourier transforms to capture long-range dependencies and extract sequence features. Despite their achievements, these methods still suffer from overfitting and transferring knowledge to new datasets. The issue lies in the lack of effective inductive biases in self-attention methods and the limitation of Fourier transforms, which only capture frequency-domain features, restricting the comprehensive modeling of users' dynamic behavior across both time and frequency domains. To tackle these issues, we propose a novel Wavelet attention-based Time-frequency domain Multi-modal Sequential Recommendation model (WTMSRec). WTMSRec consists of three core components: efficient data extraction and augmentation, a learnable frequency filter, and an innovative wavelet attention mechanism. It operates autonomously without reliance on user or item IDs, proficiently capturing and generalizing dynamic variations in user interests across both time and frequency domains. Experimental findings on five distinct downstream datasets validate that WTMSRec surpasses the performance of current cutting-edge models, with notably enhanced training efficiency.
Social networking platforms serve as prevalent mediums for social interactions, and the escalating number of users underscores the importance of categorizing account types for purposes such as platform governance, targeted advertising, recommendation systems, and more. Traditional methods typically hinge on manual rule creation to extract features, which necessitates domain expertise and poses challenges in cross-domain applicability. While deep learning methods have shown promise in representation learning, their reliance on single homogeneous information networks may hinder their ability to fully capture the intricate nature of social media, leading to potential information loss and reduced classification accuracy. To address these limitations, this paper proposes an innovative approach using an Attention-based Heterogeneous Graph Convolutional Network, dubbed AH-GCN, to learn users’ latent feature representation for social-media account classification. AH-GCN constructs a heterogeneous information network and employs random walks with restarts to sample local neighborhoods, followed by dynamic representation learning via heterogeneous graph convolution and attention mechanisms. Experimental evaluations reveal that our model significantly outperforms existing methods on two diverse social datasets, highlighting its efficacy in complex social network environments.
Error detection and accuracy optimization are two crucial aspects of program development. Various effective tools can be used to detect errors in the program and enhance computational accuracy. However, these tools can only accept expressions as input, resulting in inefficiency as they evaluate the entire program by detecting each expression sequentially. In order to solve the problem that current tools cannot directly detect the entire program, in this paper, we first propose an efficient method for floating-point (FP) program detection and optimization, which consists of two key components: 1) A novel Intermediate Representation format, FPIR, designed to uniformly represent program information and operations. 2) A novel tool, AutoFPIR, crafted to automatically convert core operations of FP program into FPIR. We then employ existing tools to detect FPIR, enabling efficient direct analysis of the entire program. Finally, we experimentally apply the proposed method to GNU Scientific Library source code programs, achieving 100% conversion accuracy and demonstrating the effectiveness of our method.
With the rapid development of AI technology, deep convolutional neural networks (CNNs) have made significant progress in real-time facial expression recognition (FER) tasks. However, traditional facial expression recognition methods still face challenges such as insufficient data diversity, high computational costs, and low recognition accuracy. To address these issues, we propose a hybrid model combining a convolutional neural network (CNN) with a graph convolutional network (GCN) for real-time facial expression recognition. By applying data augmentation techniques, we enhance the diversity of training data. The hybrid model utilizes a CNN to extract image features, which are then processed by the GCN to model the graph structure and capture the topological relationships between features. Additionally, we designed custom loss functions: geometric loss, which is based on facial key point information to enhance the geometric consistency of features, and adversarial loss, which improves the discriminative capability of the generated features. Experimental results demonstrate that the proposed model significantly improves the accuracy of real-time facial expression recognition on the CK+ dataset, while also optimizing model parameters and reducing training time. This research provides effective technical support for real-time facial expression analysis and intelligent interactive systems, showing great potential for broad applications.
In this study, we propose Qwen2-Wildfire, a lightweight multimodal model for wildfire scene recognition. The model was fine-tuned on the basis of Qwen2-VL-2B-Instruct to better recognize wildfire scenes. After fine-tuning, the model provided more detailed wildfire-related information than the pre-trained version. To customize and optimize the model, we used a high-quality dataset of 4,573 wildfire images with textual descriptions. Before fine-tuning, we optimized prompts using five multimodal labeling models: LLaVA-13b, Moondream-1.8b-v2, LLaVA-LLaMA3-8b-v1.1, MiniCPM-LLaMA3-v2.5, and LLaVA-Phi3-3.8b-mini. These models helped generate accurate text descriptions of wildfire scenes. We evaluated datasets generated by the five multimodal labeling models by using LLaVA-34b as a benchmark, comparing the similarity of the generated text to the benchmark output. A fuzzy evaluation algorithm was also used to focus on key features like flames and smoke for selecting the best-performing dataset. We then fine-tuned the Qwen2-VL-2B-Instruct model using the best-performing dataset, resulting in Qwen2-Wildfire. This study provides a high-quality wildfire dataset and the proposed Qwen2-Wildfire model for wildfire detection, achieving accurate recognition results.
