A major challenge in machine learning and data science is feature selection. Feature selection involves selecting the optimal (or suboptimal) subset of features to derive useful conclusions from a dataset based on the relevant information contained in those features. Flower Pollination Algorithm (FPA) is a metaheuristic algorithm developed recently based on flower pollination. In this paper, we propose a new type of binary FPA, called the Filter-Wrapper Modified Binary FPA (FWMBFPA), which aims to improve convergence rate and solution quality by combining filter and wrapper advantages. Using FWMBFPA, the exploration process is directed towards specific search areas by extracting the features of existing solutions. 18 UCI datasets are used to evaluate the performance of the method. FWMBFPA generally performs better than the other algorithms in terms of average classification accuracy. FWMBFPA achieves highest classification accuracy with the smallest number of selected features when compared to other algorithms when dealing with datasets with a large number of features.
This paper presents a general-purpose hardware implementation of the digital visual interface (DVI) protocol on the Xilinx Virtex-6 ML605 FPGA platform for real-time display of digital processing results. The design enables direct output of processed data from the FPGA to an external monitor without relying on external processors or software-based rendering tools. It addresses key challenges in timing synchronization, pixel formatting, and interfacing with the onboard Chrontel CH7301C encoder to support resolutions up to 1920×1080 at 60 Hz. A lightweight processing pipeline is developed in Verilog to convert multidimensional outputs into a sequential stream of pixel data conforming to the DVI protocol. As a case study, a lightweight convolutional neural network trained on the CIFAR-10 dataset is implemented on the FPGA, and its classification probabilities are displayed as a probability map on an LCD. Experimental results confirm low resource utilization and real-time performance, validating the system’s applicability in embedded applications such as machine learning inference, image processing, and real- time monitoring. This work demonstrates the feasibility of FPGA- based platforms for efficiently displaying digital video output in intelligent edge systems.
The increasing reliance on Internet of Things devices in smart grids has introduced significant cybersecurity challenges, particularly in the detection and prevention of Advanced Persistent Threats. These threats, characterized by their stealth and persistence, can compromise the integrity and functionality of critical grid infrastructure. This paper proposes the use of Deep Reinforcement Learning to enhance cybersecurity in smart grids by leveraging the ProAPT model, which is specifically designed to predict and mitigate Advanced Persistent Threats. The ProAPT model utilizes a Markov Decision Process to simulate and assess potential threats, dynamically adapting to the evolving security landscape. The model is trained using the CICAPT-IIoT dataset, which includes simulated attack scenarios in industrial IoT networks. The results of our experiments demonstrate the effectiveness of the ProAPT model in detecting and preventing APTs in smart grid environments. Experimental results show that the ProAPT model significantly outperforms traditional machine learning algorithms like Random Forest, Support Vector Machines, and Logistic Regression, achieving 93.8% accuracy, 93.12% precision, 95.2% recall, and 94.15% F1-Score. The feature importance analysis reveals that traffic-related features such as packet size variance and connection duration are crucial in identifying Advanced Persistent Threats. This paper demonstrates the effectiveness of Deep Reinforcement Learning in enhancing smart grid cybersecurity by proactively identifying and mitigating cyber threats, offering a promising approach to securing IoT-based critical infrastructures against sophisticated cyberattacks.
This paper investigates the deployment of over-the-air federated learning (OTA-FL), leveraging the dynamic repositioning and line-of-sight communication capabilities of unmanned aerial vehicles (UAVs) and movable antennas to enhance network efficiency. A closed-form expression is derived to quantify the optimality gap between the actual federated learning (FL) model and its theoretical ideal, accounting for the capabilities of movable antennas to show the diverse relationship between Mean Square Error (MSE) and the optimality gap. Then An MSE minimization problem is then formulated, involving the joint optimization of moveable antenna position vectors, and the beamforming vector at the UAV. This complex non-convex problem is reformulated as a Markov Decision Process (MDP) and solved using the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm within the deep reinforcement learning (DRL) framework. Numerical results demonstrate that the proposed algorithm outperforms benchmarks such as Advantage Actor-Critic(A2C) and Soft Actor-Critic (SAC).
Deep neural networks typically require predefined architectures, which can lead to overfitting, underfitting, high computational costs, and storage overhead. Dynamic structure optimization through pruning can reduce network redundancy, but it often results in performance degradation. In this study, we propose a novel pruning method inspired by biological synaptic pruning that adaptively optimizes deep neural network structures. The proposed method continuously monitors the contribution of each connection during training using a dynamic efficiency criterion that evaluates the relative importance of each connection within its layer. Connections are not removed immediately; only those consistently falling below a predefined threshold are pruned, ensuring stability and robustness. Validation is conducted by simulation on an industrial distillation column dataset under noisy conditions and the MNIST benchmark dataset. The results demonstrate improved accuracy, enhanced generalization, and faster learning, with an average pruning rate of 53%. Compared to conventional and state-of-the-art pruning techniques, our method achieves superior performance in terms of compression rate and accuracy while effectively mitigating overfitting.
Influence Maximization (IM) is a fundamental problem in social network analysis that seeks to identify a small set of highly influential nodes that can maximize the spread of information. Due to its NP-hard nature, finding an exact solution is computationally infeasible for large-scale networks. To address this, this paper introduces an enhanced discrete Manta-Ray Foraging Optimization (MRFO) algorithm tailored for IM. The proposed method integrates degree, closeness, and betweenness centrality measures into the fitness function and introduces a fused centrality index to improve the identification of influential nodes. To handle the discrete search space, the continuous MRFO is adapted with novel discretization mechanisms. Experimental evaluations on five real-world networks (NetScience, Email, Hamsterster, Ego-Facebook, and Pages-PublicFigure) demonstrate that the proposed method achieves higher influence spread compared to existing baseline algorithms, with average improvements of 14.63%, 12.81%, 19.03%, 15.24%, and 18.76%, respectively. These results validate the effectiveness, robustness, and practical applicability of the proposed approach for large-scale IM.
Since the genesis of layered network, designing a popper MAC control protocol was a major concern. Among many protocols which introduced earlier, there is always a trade-off between utilization and load overhead. ALOHA is one of the first MAC protocols with virtually possess no overhead, but its maximum throughput is limited. Hence a new MAC protocol introduced on basis of multi-packet reception model named Hybrid ALOHA. In the original paper stability and throughput of this algorithm for 2 or 3 users case system had been analyzed. Although stability region for above two users circumstances had been studied, there was no general form for throughput nor any practical examination of stability. In this paper, beside expanding formula for throughput for any arbitrary number of users, the throughput of system is checked with simple simulation of probability of successes and failures. Achieved results shows that regardless of additional overhead for more users, throughput remains proper, and the system is not lost stability in larger number of users.
It is concerning that the growing popularity of social networks is encouraging violence or inciting offense toward other people. An attempt has been made in the past several years to detect offensive language in social media posts. Nonetheless, the majority of studies focus on recognizing offensive language in English. Moreover, dataset labeling emerges as a crucial and fundamental step for training high-quality models, considering the increasing use of artificial intelligence and machine learning tools. Utilizing crowdsourcing platforms is an efficient and optimal method that can be used for data labeling. This approach uses human resources who are sufficiently knowledgeable about the topic to label the data. In this paper, we introduce PerGOLD, a new Persian General Offensive Language Dataset, in which we use an event-based data collection methodology to detect offensive language in Persian Twitter. To access labeled training data, we build a crowdsourcing platform to benefit from human input. We labeled 13,716 tweets, and according to the obtained results, 34% of them were labeled as offensive language. Finally, we evaluated the efficiency of these data by applying some classic machine learning models (LR, SVM) and transformer-based language models (RoBERTa, ParsBERT). The obtained F1-score of the best model (ParsBERT) was 85.4%.
COVID-19 has created an urgent need for innovative detection methods. This study presents a novel approach to identifying potential COVID-19 patients by analyzing their WhatsApp messages using advanced natural language processing techniques. Our methodology combines Word2Vec embeddings with lexical-semantic enrichment using ConceptNet, creating a comprehensive system that can detect subtle linguistic patterns associated with COVID-19 symptoms and experiences. The system processes WhatsApp messages through multiple stages: initial data collection, Word2Vec embedding, lexicon semantic enhancement, vector-space model creation, Biterm Topic Model-based feature selection, and finally, Naive Bayes classification. By enriching the language model with synonyms and capturing complex semantic relationships, our approach can identify potential COVID-19 cases based on how people describe their symptoms and experiences in everyday conversations. We tested the system on a sample of diverse WhatsApp messages, achieving promising results in distinguishing between messages from COVID-19 patients and healthy individuals. The system successfully identified both explicit statements of COVID-19 status and more subtle descriptions of symptoms, while correctly classifying non-COVID related messages with high confidence. While this method shows potential as a non-invasive and scalable screening tool, it should be viewed as complementary to existing diagnostic approaches rather than a replacement. Further large-scale testing is needed to fully validate the system's reliability and effectiveness in real-world applications.
This paper presents a model-driven engineering framework designed to enhance the development of flexible, high-quality audio-based applications on mobile platforms. The framework comprises domain-specific metamodels, a graphical editor, and a transformation engine, enabling the automatic generation of application code and supporting customization within Android Studio. To address the challenges faced by developers in delivering effective audio applications, the framework provides a structured approach to simplify design and implementation processes. The framework’s applicability is demonstrated through four case studies, highlighting its ability to create diverse audio-based Android applications. A detailed evaluation includes a comparison of development effort between the proposed model-driven approach and traditional coding methods, showing significant reductions in time and manual effort. Additionally, the framework is assessed using key software quality metrics such as maintainability, understandability, and extensibility. The findings demonstrate that the model-driven approach not only streamlines development but also improves the maintenance of applications, enabling developers to meet the growing demand for audio applications efficiently. By reducing development costs and enhancing productivity, this research contributes to the field of software engineering, offering a practical and adaptable methodology for audio-based application development.
Scene understanding through semantic segmentation is a vital component for autonomous vehicles. Given the importance of safety in autonomous driving, existing methods are constantly striving to improve accuracy and reduce error. RGB-based semantic segmentation models typically underperform due to information loss in challenging situations such as lighting variations and limitations in distinguishing occluded objects of similar appearance. Therefore, recent studies have developed RGB-D semantic segmentation methods by employing attention-based fusion modules. Existing fusion modules typically combine cross-modal features by focusing on each modality independently, which limits their ability to capture the complementary nature of modalities. To address this issue, we propose a simple yet effective module called the Discriminative Cross-modal Attention Fusion (DCMAF) module. Specifically, the proposed module performs cross-modal discrimination using element-wise subtraction in an attention-based approach. By integrating the DCMAF module with efficient channel- and spatial-wise attention modules, we introduce the Discriminative Cross-modal Network (DCMNet), a scale- and appearance-invariant model. Extensive experiments demonstrate significant improvements, particularly in predicting small and fine objects, achieving an mIoU of 77.39% on the CamVid dataset, outperforming state-of-the-art RGB-based methods, and a remarkable mIoU of 82.8% on the Cityscapes dataset. As the CamVid dataset lacks depth information, we employ the DPT monocular depth estimation model to generate depth images.
The Internet of Vehicles (IoV) represents a transformative paradigm in Intelligent Transportation Systems (ITS), enabling real-time communication between vehicles, infrastructure, and cloud platforms to improve traffic management, safety, and efficiency. However, the resource limitations in vehicles pose significant challenges for delay-sensitive applications such as autonomous driving and automated navigation. Vehicular Edge Computing (VEC) offers a promising solution by offloading tasks to edge servers near vehicles, reducing transmission delays and enhancing computational efficiency. In this paper, we address the complex task offloading and resource allocation problem in VEC environments. We model this challenge as an Integer Linear Programming (ILP) problem, aiming to maximize the system’s overall profit. To mitigate the computational complexity of solving the ILP problem, we propose an efficient heuristic algorithm. This approach considers various task types, accounting for the diversity and specific requirements of each. The algorithm optimizes CPU resource allocation based on task generation rates, average task sizes, and a calculated weight coefficient for each task type. Simulation results demonstrate that the proposed algorithm reduces memory costs and penalties from rejected tasks, while improving overall system profit. In particular, it outperforms existing algorithms by an average of 18.26% in terms of profit, demonstrating its effectiveness in practical VEC applications.
Twitter List recommender systems can generate highly accurate recommendations, but since they employ heterogeneous information of users and Lists and apply complex prediction models, they cannot provide easy understandable intrinsic explanations. To address this limitation, Twitter List descriptions can play a critical role in providing post-hoc explanations that help users make informed decisions. In this paper, we propose an explanation model to provide relevant and informative explanations for recommended Lists by automatically generating descriptions for Twitter Lists. The model selects the most informative tweets from a List as its description to inform users more with the recommended List that positively contributes to the user experience. More specifically, the explanation model incorporates three categories of features: content relevance features, tweet-specific features, and publisher’s authority features that are used in a learning to rank model to rank the List’s tweets in terms of their informativeness. By conducting experiments on a Twitter dataset, we have shown that the proposed model provides useful explanations for the Lists that are recommended to users, while upholding parity in recommendation performance.
In today's era, the Internet of Things has become one of the important pillars in organizations, hospitals, and research circles and is recognized as an integral part of the Internet. One of the important areas that require online monitoring is medical imaging equipment, whose functional information is transmitted through the Internet of Things. Server security and intrusion prevention, along with anomaly detection, are critical requirements for these networks. The purpose of anomaly detection is to develop methods that can detect attackers' attacks and prevent them from happening again. Algorithms and methods based on statistics play an important role in predicting and diagnosing anomalies. In this article, the isolation forest algorithm was used for training on 80% of the dataset related to the data of the Internet of Medical Things network, and then this model was tested and evaluated on the remaining 20%. The results show 90.54% accuracy in detecting anomalies in the received data, which confirms the effective performance of this method in this field.
Abstract-- In the domain of software development, the evaluation of developer expertise has gained prominence, particularly with the rise of serverless functions. These functions, which simplify the development process by delegating infrastructure management to cloud providers, are becoming more common. As developers may utilize functions created by their peers, understanding the expertise of the original developer is crucial since it can serve as an indicator of the functions' quality. While there are existing methods for expertise evaluation, certain gaps remain, especially concerning serverless functions. To address this, our research aims to enhance the assessment of developer expertise in this area by extracting activity-based features from both GitHub and Stack Overflow. After processing the extracted data, we applied various machine learning algorithms. Our findings suggest a potential improvement in evaluating developer expertise when incorporating features from Stack Overflow compared to using only GitHub data. The extent of this improvement was observed to differ among programming languages, with variations in accuracy improvement percentages ranging from 2% to 19%. This study contributes to the ongoing discourse on developer expertise evaluation, highlighting the potential benefits of drawing from multiple data sources.
The increasing popularity of vehicular communication systems necessitates efficient and autonomous decision-making to address the challenges of vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communications. In this paper, we present a comprehensive study on channelization in Cellular Vehicle-to-Everything (C-V2X) communication and propose a novel two-layer multi-agent approach that integrates deep reinforcement learning (DRL) and federated learning (FL) to enhance the decision-making process in channel utilization.Our approach leverages the autonomy of each vehicle, treating it as an independent agent capable of making channel selection decisions based on its local observations in its own cluster. Simultaneously, a centralized architecture coordinates nearby vehicles to optimize overall system performance. The DRL-based decision-making model considers crucial factors, such as instantaneous channel state information and historical link selections, to dynamically allocate channels and transmission power, leading to improved system efficiency.By incorporating federated learning, we enable knowledge sharing and synchronization among the decentralized vehicular agents. This collaborative approach harnesses the collective intelligence of the network, empowering each agent to gain insights into the broader network dynamics beyond its limited observations. The results of our extensive simulations demonstrate the superiority of the proposed approach over existing methods, as it achieves higher data rates, success rates, and superior interference mitigation.
For computer networks to remain secure, intrusion detection is essential. Analyzing network traffic data is part of this activity to spot possible cyber threats. However, the curse of dimensionality presents a challenge because there are so many dimensions in the data. To overcome this challenge, feature selection is essential to creating a successful intrusion detection system. It involves removing irrelevant and redundant features, which enhances the classification model's accuracy and lowers the dimensionality of the feature space. Metaheuristic algorithms are optimization techniques inspired by nature and are well-suited to choose features for network intrusion detection. They are effective in exploring large search spaces and have been widely used for this purpose. In this study, we improve the Sine Cosine Algorithm named ISCA for feature selection by introducing a controlling parameter to balance exploration and exploitation. Based on the NSL-KDD dataset, the results show that compared to other competing algorithms, the ISCA performs better than other metaheuristic algorithms in terms of both the number of features selected and the accuracy of classification.
The COVID-19 pandemic has highlighted the urgent need for rapid and accurate diagnostic methods. In this study, we evaluate three machine learning models—Random Forest (RF), Logistic Regression (LR) and Decision Tree (DT)—for detecting COVID-19 trained on preprocessed imbalanced datasets with 5086 negative and 558 positive cases. To this end, we demonstrate the capability of two advanced data synthesis algorithms, Conditional Tabular Generative Adversarial Network (CTGAN) and Tabular Variational Autoencoder (TVAE), in addressing the class imbalance inherent in the dataset. The classifiers trained on the original as well as the balanced datasets were evaluated for comparison. Our findings reveal that RF obtains the highest accuracy of 98.83% on the CTGAN-balanced dataset. In conclusion, our results verify the potential of coupling data synthesis with traditional machine learning for the diagnosis of COVID-19. We hope that this research will make a significant contribution to the current AI (Artificial Intelligence) efforts in combating the pandemic.
With the growing use of social media, figurative language has become very common on social media platforms. Given its complexity, figurative language can confuse natural language processing systems and lead to incorrect results. To address this issue, researchers have developed methods to detect humor, jokes, irony, and especially sarcasm. To date, most studies have used deep learning methods to identify sarcasm. Some studies have also incorporated context such as previous posts or conversations to improve the accuracy of sarcasm detection. But the context that can be highly effective in detecting the sarcasm of posts is the characteristics of the writer of the posts. So, the present paper aims to develop a hybrid approach that combines content and context features to better identify sarcastic posts. i.e., this study additionally proposes a deep learning method to model the content of tweets and suggests a multi-dimensional method that considers the user’s writing style and personality traits as context features. Several experiments were used to evaluate the effectiveness of the proposed method. The results indicated that the proposed method outperformed baseline methods in sarcasm detection.
Nowadays, the number of cyberattacks and, consequently, the risk of them is increasing significantly. So, it is vital to be aware of the cybersecurity situation. This paper provides a model to measure the success or effectiveness of the organization's security missions, that is, to determine whether security events such as the occurrence of an attack or the selection of countermeasures have been effective on the success of organizational missions. Two components are considered as inputs for this purpose. The first one is the network dependency graph which demonstrates how different assets in the network are dependent on each other and with what intensity or weight they affect each other. Another component is the mission dependency graph which specifies the relation between organizational assets, tasks, functions, and mission objectives. It also specifies the impact of the assets on the organizational tasks, functions, and mission objectives. Previous researches focused more on determining the impact of attacks on different assets. However, this paper aims to assess this impact by considering the organizational mission. This model is proposed in such a way that its components are separable. It helps different organizations with specific goals and different requirements to use this model so that they can personalize and customize its different components. This model can be employed for critical asset recognition. Moreover, it enables us to know which countermeasures are more effective in terms of the organizational mission.
