نشریه محاسبات نرم
سال چهاردهم شماره 2 (پیاپی 28، پاییز و زمستان 1404)

Knowledge graphs have emerged as vital tools extensively used in fields such as data mining, machine learning, complex networks, and related domains. This paper aims to conduct a comprehensive review of these graphs and the methods associated with them. Specifically, we analyze and evaluate the RGNN, TDM, and BLM network methods in more detail, carefully reviewing the results obtained from each. This study provides a comprehensive analysis of the performance of these methods and evaluates their strengths and weaknesses in real-world challenges and applications. We hope that this review will contribute to a better understanding of this important technology and facilitate its future development.

This article discusses the use of the whale optimization algorithm and the integral time-absolute error criterion (ITAE) to adjust the optimal parameters of the backstepping controller. The goal is to improve the performance of active power filters based on LCL filters in four-wire distribution systems. By using a shunt active power filter with the proposed method at the point of common coupling (PCC), the approach minimizes harmonics in the grid current and makes it more sinusoidal. In most references, the selection of backstepping controller coefficients is based on trial and error, which, in addition to being time-consuming, provides no guarantee for the correctness of the selection. The purpose of this article is to find the sub-optimal parameters of the backstepping controller using the whale optimization algorithm, known for its high accuracy and convergence speed. By minimizing the selected error criterion, the algorithm sub-optimally determines the backstepping parameters, significantly improving the controller's performance and ensuring the proper filtering of grid current harmonics. The effectiveness of the proposed method is demonstrated through simulations of different scenarios. The results show that the proposed method performs very well in improving power quality in distribution systems.

Vehicular Ad-hoc Networks are a growing technology that ensures safe and efficient road traffic management. Mobility management is the primary challenge in inter-vehicle networks. Therefore, an adaptive and stable clustering protocol is a crucial issue in inter-vehicle network technology. Day by day, inter-vehicle networks based on clustering are attracting more attention. The main variables used in forming clustering protocols are speed, distance, and direction of movement. However, in different methods, other variables such as confidence level, lifespan, and similarity measure, quality of communication link, signal noise rate, and packet transmission rate have been used alongside the three main components. In this paper, we aim to improve aspects such as the number of clusters formed and the number of changes in cluster heads by employing various methods and techniques presented. The purpose of these improvements is to create more cohesion and stability in the formed clusters. In the proposed algorithm, we aim to improve the number of clusters formed by at least 5.7% by using spatial indexing and examining the linear position that the vehicle passes through. Additionally, we have achieved a 3.2% improvement in the number of cluster head changes.

Requirements engineering is a critical phase within the software development lifecycle, with its effective execution being pivotal to the overall success of a project. To improve the quality and efficiency of requirements engineering, the reuse of previous requirements has emerged as a promising strategy. By employing methodologies such as collaborative filtering, content-based filtering, knowledge-based techniques, and hybrid approaches, recommender systems support stakeholders in the precise identification and prioritization of requirements. This paper presents a hybrid method combining collaborative filtering and content-based filtering to improve the performance of recommendation systems in requirements engineering. In this method, the semantic similarity between requirements is calculated using the Global Vectors for Word Representation (GloVe) model, and the semantic features extracted from textual descriptions are combined with predefined categories. This process enhances the accuracy of predicting stakeholder interest or need for requirements and their prioritization. The experimental results obtained from the RALIC dataset indicate that the proposed approach significantly enhances prediction accuracy and coverage. This improvement is particularly evident in addressing challenges related to data sparsity and the cold start problem, highlighting the superior efficiency of this method compared to existing approaches.

Nowadays, many researchers use systems based on deep learning and artificial intelligence to diagnose and classify many diseases in the field of medicine, such as cancer diagnosis. Early and accurate diagnosis of breast cancer based on mammography images can greatly help doctors in timely treatment of the disease. On the other hand, most of the deep networks used in breast cancer (BC) detection systems are not suitable for real-time applications due to the use of fully connected (FC) layers, time-consuming training method, and high calculation volume. Therefore, in this article, a new hybrid method based on deep learning and swarm intelligence called EZFNet-ELM-COA is proposed for real-time detection of breast cancer. In the proposed network (Enhanced ZFNet), in the convolutional layers of the ZFNet network, large size filters have been replaced by two series of smaller size filters, and then the outputs of the filters have been joined together. The use of this multi-scale technique makes the network extract more effective features. Also, in order to increase the speed of training and significantly reduce the volume of calculations, fully connected layers of the pre-trained EZFNet network were replaced by an Extreme Learning Machine (ELM). At the end, the generalization capability of ELM has been improved by updating the weights and biases of the ELM layer by the Coati Optimization Algorithm (COA). In order to check the performance of the system, six evaluation criteria have been used, and our proposed model achieved values of  97.97, 95.70, 97.15, 98.28, 96.42 and 93.98 percent for accuracy, precision, sensitivity, recognizability, F1-score and MCC, ​​respectively. The obtained results show the increase of accuracy, improvement of efficiency, significant reduction of calculation time, and high performance level of the proposed system compared to the existing methods.

With the evolution of the fifth generation wireless network, the Internet of Things (IoT) has also created a surprising transformation in human life. Therefore, a more detailed study of IoT technology and its various fields, including security, is necessary. In the field of IoT network security, one of the issues that has been raised is the spread of malware. In this paper, we have modeled the spread of malware in IoT networks with the epidemic disease model of vulnerable - exposed to infection - infected - recovered - antidote. The effect of applying antidote to recovered and vulnerable nodes on the spread of malware in IoT networks has been investigated. We have determined the importance of the node in the network based on its degree, and based on this rate, the vulnerable nodes in the network have been secured. Securing nodes based on the degree has led to a reduction in the spread of malware in the network. The epidemic threshold of malware spread (R0) has been calculated in the proposed model. Comparing the model with the SIRS and SEIRS models showed that the proposed model performed better than the other two models.

Underwater wireless sensor networks (UWSNs) face various constraints such as limited node energy, sensor mobility at different depths, and the need for diverse communication methods. These limitations reduce the efficiency of conventional routing algorithms used in other multi-hop wireless networks. In this study, an innovative routing method based on reinforcement learning and generative adversarial neural networks (GANs) is proposed for UWSNs. The proposed approach aims to discover and store optimal data transmission paths within the network. Subsequently, these paths are used to train a deep learning model following the generative adversarial neural network framework, allowing for the generation of new routes. Simulation results demonstrate that the proposed method improves the successful packet delivery rate and extends network lifetime compared to traditional reinforcement learning techniques in UWSNs.

In this paper, an efficient conjugate gradient method is introduced for solving unconstrained nonsmooth optimization problems. Conjugate gradient methods are among the most popular methods for solving smooth optimization problems due to their simplicity and low memory requirements; however, their application to nonsmooth problems has received less attention. Therefore, the Lipschitz continuous objective function is first smoothed using the Moreau–Yosida regularization function. A new descent direction is then proposed by combining the first-order derivative information of the smoothed function with the previous descent direction. Using an inexact line search technique, we show that the generated direction satisfies the sufficient decrease condition and that the new iterate remains within a suitable trust region relative to the previous iterate. Moreover, the global convergence of the proposed method is guaranteed under standard assumptions. Finally, the effectiveness of this method is evaluated in the field of image recovery, and the results demonstrate its superior performance compared to existing methods.

The success of students at every educational level is one of the significant goals of the Ministry of Education. In the meantime, proper planning for the academic success of tenth-grade students is more valuable because of the choice of study field in this grade. Likewise, making a correct decision requires gathering comprehensive information from all aspects of a student's life. Therefore, the life balance wheel tool in six dimensions -physical, financial, intellectual, emotional, social, and spiritual- can be a suitable solution. In this article, first, a dataset is created using the life balance wheel, along with demographic and educational background information of 10th-grade students from educational institutions in Kashmar, collected through a questionnaire. Then, well-known Wrapper and Filter methods are employed for selection of affective features and data mining algorithms including Multilayer Perceptron, J48 decision tree, Random Forest, Naive Bayes, SVM, and KNN are used to predict the success of students in 10th grade. The results show that the Wrapper algorithm can select more impact features than others; specifically, the features of pocket money, study time, gender, field of study, father’s education, mother’s education, and spiritual dimension have the most impact. Also, the Multilayer Perceptron algorithm with 85.62% accuracy and the Naïve Bayes algorithm with 86.25% accuracy demonstrated the best performance before and after the feature selection process, respectively.

Capital accumulation and proper asset management play a crucial role in the economic growth of countries. The significant impact of this factor, as well as its importance, can be clearly observed in the system of most countries, especially capitalist systems. The capital market is one of the most appropriate places to attract small assets and use them in order to grow and improve a company and also increase the personal assets of investors. Therefore, correct investment helps individuals to manage the surplus of their incomes properly, enabling them to attain the necessary capital to achieve their goals. Essentially, stock prices are nonlinear and chaotic; hence, investing in the stock market carries a high risk. To minimize this risk and reduce hazards, an efficient system is required that can predict future stock price movements with high accuracy. Given the importance of this issue, machine learning models have shown suitable performance in this area. Machine learning algorithms have the ability to model nonlinear and complex systems. In this research, a novel collective approach based on Support Vector Regression is used to construct a composite or hybrid model that enjoys high accuracy and speed in predicting stock prices. The stocks used in this research are the ten index-making shares of FOLD, TAMN, PKLJ, MSMI, PNES, PTEH, PNBA, IKCO, SIPA and PARS, considered in the period from 2021-03-27 to 2024-07-21. Based on various criteria, the results demonstrate the superiority of the collective Support Vector Regression method over the conventional Support Vector Regression method and the Random Forest approach.

COVID-19, mainly known as Corona, is a viral disease caused by the SARS-CoV-2 virus. The symptoms of COVID-19 include fever, fatigue, cough, and a loss of the sense of smell. The rRT-PCR test, as the standard diagnostic tool for this disease, requires respiratory sampling, which is time-consuming. Therefore, developing rapid diagnostic methods is of great importance. In this paper, a novel, rapid, and non-invasive method for diagnosing COVID-19 using artificial intelligence is introduced. This method comprises two stages: feature selection and disease diagnosis, which are performed using the Tree Social Relationships (TSR) algorithm and the Naive Bayes classifier. The proposed method is practically evaluated using the COVID-19 Dataset. Experimental evaluations show that this method outperforms existing approaches in diagnosing COVID-19, achieving 96% accuracy, 97% recall, and a 96% F1-score.

Sentiment analysis is one of the important branches of natural language processing, which aims to classify texts with respect to the feelings and attitudes of the author of the text. In Persian, most of the available sentiment lexicons are small in size and lack slang expressions and informal words. These features significantly reduce the performance of sentiment analysis algorithms. This paper aims to present a method based on deep learning and sentiment lexicons for sentiment analysis of Persian texts written on social networks. Since most existing sentiment lexicons in Persian language are small in size and lack slang and informal expressions, first, two methods based on ChatGPT are proposed to expand the existing Persian sentiment lexicons by adding slang expressions that are widely used in social media. Then, the combination of the sentiment lexicon and dual-channel convolutional neural network (DC-CNN) is used to determine the polarity of texts. Experimental results show that by expanding the existing sentiment lexicons with the two proposed methods, the accuracy of the sentiment analysis algorithm increases by 1.74 and 2.14 percent, respectively, which indicates the success of ChatGPT in polarity classification of Persian slang expressions. Also, employing the features extracted from the sentiment lexicon in a DC-CNN leads to an increase in the precision of the two base models by 1.6 and 3.2 percent.