نشریه محاسبات نرم
سال چهاردهم شماره 1 (پیاپی 27، بهار و تابستان 1404)

Nowadays social networks have emerged as a popular platform for communication, entertainment, news and event updates, and more. While social networks offer several benefits such as e-commerce and easy communication, the presence of rumors is a significant drawback. Rumors can spread quickly to many users with minimal effort and lead to unpleasant social phenomena. In previous works, researchers have proposed machine learning and deep learning approaches for automatic rumor detection. This paper proposes a circular multi-layered CNN-BILSTM model for rumor detection in social networks. The model transforms the multi CNN-BILSTMs into a convolutional matrix and then applies a circular state to the matrix. The model only uses the text of social network users and classifies it as rumor or non-rumor. We tested the model on the real public dataset of PHEME and Liar. The model outperforms twelve state-of-the-art models in terms of accuracy.

Link prediction is one of the important topics in network analysis. Link prediction can be done by a classifier such that the feature vector of a pair of nodes is its input. The output of the classifier indicates whether a link is predicted between that pair of nodes (class one or class zero). To extract the feature vector of a pair of nodes, neural networks (GNN) can be used, in which case the method of solving the link prediction problem will be based on GNN. In this paper, a GNN-based link prediction problem solving method called GAE is considered as the basic method. One of the basic problems in this method is that the feature vector extracted by GNN can be the same for different pairs of nodes. To solve this problem, in this paper, using the concept of subgraph, the basic method is improved and a new framework called SGAE is proposed. The proposed framework has been compared with the basic methods based on different evaluation criteria, which shows its improved performance. For example, the SGAE method has improved 5.5, 5, 5.75 and 5.87 compared to the basic GAE method in terms of accuracy, F1-Score, average precision and area under the precision-recall curve.

A simple mathematical model of human liver function is a linear system of differential equationsconsisting of two unknown functions and their first order derivative.Since laboratory data and clinicalvalues are recorded by measurement and evaluation, they have the characteristic of uncertainty andinaccuracy. This paper deals with the analysis and solving of the model with fuzzy data. At first, theproblem is investigated in fully fuzzy case from the point of view of generalized Hukuhara differentiability,and sufficient conditions for the existence of a unique solution are found. Next, the simplerforms of the problem are studied in two separate cases of the fuzziness of parameters, and in eachcase, the solution formulas are obtained, in details. In the end, two numerical examples are solved todescribe and apply the results in practice.

In this paper, a hybrid method is presented to solve the fuzzy time-fractional Klein-Gordan equation. This method is a combination of finite difference and fractional differential transform methods. In this way, using the semi-discretization technique, the lower branch of the fuzzy Klein-Gordan equation is converted to a system of ordinary differential equations. Then, the obtained system of differential equations is solved by the fractional differential transform method, and the solution of the lower branch is obtained. Finally, by using the same procedure for the upper branch, the approximate solution of the upper branch, and the approximate solution of the fuzzy fractional Klein-Gordan equation is obtained. The convergence is examined, also, the efficiency and accuracy of this method are approved by solving some examples using the proposed method. Also, the CPU time shows the computational cost of this algorithm is low.

Due to the increasing activities of individuals and legal entities in the capital market and the transformation of this market into one of the most important economic drivers of any country, it is concluded that more knowledge regarding the selection of shares will undoubtedly lead to higher profitability. In this paper, clustering of stock price time series using the smallest spanning tree algorithm is suggested. The daily closing prices of the shares of the companies listed on the Tehran Stock Exchange from 09/23/2019 to 11/10/2020 are used as the dataset. In the first stage, we form some sub-clusters that include similar companies in terms of price behavior. Then, based on a similarity criterion, sub-clusters are merged until the desired clusters, which contain members with the most similarity, are achieved. The main advantage of the proposed method is that the similarity measures are calculated locally, resulting in lower computational costs compared to other methods. The results indicate that the method can easily perform the clustering process, especially for large datasets, with favorable accuracy.

Performance predicting of GPUs is a challenge for CUDA programmers and has a wide application in the tuning of these programs. This is because the overheads of CUDA programs may make running programs on a GPU uneconomical. Performance prediction helps the programmer to tune the programs more consciously and find the optimal point of the tuning more accurately instead of trial and error. Having a model that can predict performance can help programmers in manual tuning and source-to-source compiler developers in auto-tuning CUDA programs. In this paper, we presented a performance model for predicting the execution time of a CUDA kernel. In this model, first by using profiling and by using the static analysis of CUDA and PTX codes, two datasets have been created for the execution time of computation-bound and memory-bound kernels. Then, using the methods of the artificial neural network, support vector machine, and extreme learning machine, the execution time of the CUDA kernel is predicted. The experimental results show that the extreme learning machine method can predict the execution time of a computation-bound kernel with a maximum error of 3.42% and the execution time of a memory- bound kernel with a maximum error of 9.84%. The cross-validation method was used to validate the results. Also, by using feature selection methods, the impact of each input feature on kernel execution time has been determined.

In this paper, we present a finite element method to approximate the solution of optimal control problems involving distributed-order time-fractional diffusion equations. The dynamics of these problems involve distributed-order time-fractional derivatives, which are a generalization of fractional derivatives. Despite the importance of these problems, there exist few researches on their solving in the literatures. Numerical methods for solving optimal control problems are classified into two main categories, namely, indirect methods and direct methods. In the indirect methods, by using Pontryagin principle, the necessary conditions for optimality are derived and formulated as a nonlinear two-point boundary value problem. On the other hand, in the direct methods, by discretizing the control and state variables, the considered problem is reduced to a nonlinear programming problem. Due to the difficulties related to solving the system of equations resulting from the necessary conditions for optimality in optimal control problems involving distributed-order time-fractional diffusion equations, in this paper we use the approach of direct methods to approximate the solution of these problems. To approximate the distributed-order time-fractional derivatives, we employ the Grunwald-Letnikov and L1 approximation methods and derive two approximation formulas for derivative. Also for spatial discretization, we utilize the piecewise linear finite element method. Therefore, we transform the original problem into a convex quadratic optimization problem, which can be efficiently solved using existing optimization algorithms. We consider two numerical examples to demonstrate the efficiency and accuracy of the proposed method.

The users’ interactions in social networks can reflect their personality traits. Therefore, by analyzing the data of users' interactions, their personality traits can be estimated. Then, the estimated personality traits can be used for various purposes, including monitoring the mental health of users, personalizing content and services to the users, marketing, and advertising activities. In this study, the focus is on the self-esteem personality trait, and models are presented to estimate the self-esteem of Iranian users of the Instagram social network. For this purpose, first, a population of Instagram users was considered. These users completed the Rosenberg questionnaire and their self-esteem was measured. Then, various features (e.g., the number of followers and followers of users, and their profile image specifications) were also extracted from the account of these users to be used together with their self-esteem level. By applying machine learning algorithms to these data, models were extracted to estimate users' confidence. The evaluation shows that the proposed method can estimate users' self-esteem based on the information of users’ Instagram accounts with the accuracy of 0.81 and the precision of 0.77. Also, the results showed that if the self-esteem of males and females is modeled separately, more accurate models will be obtained.

Nowadays, fault tolerance in different systems is a very essential factor. Using checkpointing methods and safe spots for recovery after faults occur can increase the reliability and dependability of systems. The main issue with using checkpointing methods is their overhead. This overhead made as a result of checkpointing execution and it has negative impact on system performance. Therefore, numerous approaches and methods have been introduced to address this problem. These approaches and methods aim to reduce the overhead in order to increase system performance. this paper, thoroughly studied and reviewed various checkpointing methods. These methods organized into distinct groups. Then, determine These groups based on the type of checkpointing execution and the different systems levels. Those are such as: coordinated checkpointing, system-level checkpointing, application-level checkpointing, and distributed system checkpointing. Finally, this paper provides a detailed summary in a Comprehensive graph and conclusion for each of these groups.

Aspect-based sentiment analysis (ABSA) of tweets using deep neural networks has garnered significant attention. However, these networks generally perform best with access to extensive, high-quality data. This research leverages zero-shot deep transfer learning and cross-domain data to enrich the educational resources available within the target domain. The study began by gathering tweets from cryptocurrency influencers, followed by employing a combination of a deep neural network, an attention mechanism, and the pre-trained DistillBERT model for aspect-based sentiment analysis. The attention layer's heat map indicates that integrating this layer with deep models enhances the identification of aspect-specific words. In zero-shot transfer learning, the model operates without labelled target data samples, relying exclusively on cross-domain data for training. Additionally, using cross-domain data with optimal selection based on Pearson's similarity coefficient helps minimize negative transfer under low-data conditions, improving the model's sensitivity to textual context. To address data imbalance, the Synthetic Minority Over-Sampling Technique (SMOTE) was applied. Experimental results indicate a 2% improvement in average accuracy and F1 score on the SemEval benchmark dataset, with the ROC-AUC reaching 86.35% relative to previous studies. Furthermore, we employed an adaptive learning rate to expedite model convergence and used Grid Search to optimize the network's hyperparameters, enhancing its adaptability to the target domain.

In the field of numerical analysis, one of the most difficult problems to solve is nonlinear equations. Traditional numerical techniques like Newton's methods and their variations typically demand a well-informed initial guess to successfully solve these nonlinear systems. An improper initial guess can significantly hinder the performance and convergence of these methods, making it difficult and time-consuming to achieve the desired results in practice. To overcome these issues, this paper presents a novel approach that employed the Improved ARO (IARO) meta-heuristic algorithm for solving nonlinear equation systems numerically. Given that solving nonlinear equation systems can be transformed into an optimization problem, meta-heuristic algorithms exhibit strong potential for finding solutions efficiently. The ARO meta-heuristic algorithm, inspired by the feeding behavior of rabbits, has demonstrated its ability to solve intricate optimization problems within acceptable time. In the proposed method, the ARO algorithm is improved with adding a memory table to have a suitable performance for solving nonlinear equation systems. To assess the effectiveness of the proposed method, it is applied to solve several complex nonlinear equations. The results clearly demonstrate the robust performance of the approach.

Diabetes is one of the most significant factors leading to death, which can significantly result in kidney diseases, heart diseases, and sight loss. The application of data mining could be helpful for the diagnosis and treatment of this disease. Predicting diabetes is an important area of research that can help improve the treatment process. It is necessary to prevent, monitor, and raise awareness about this disease. In this article, a new method to diagnose diabetes is proposed. The proposed method includes a pre-processing stage in which outlier data is removed. Eventually, by using the K-Nearest Neighbor classifier, Extreme Gradient Boosting and CatBoost, samples have been classified into two classes: diabetic and non-diabetic. In the end, to improve the proposed method, a soft voting algorithm has been used to merge the three classifiers. The proposed method has been applied to the Pima diabetes dataset, which includes information on age, gender, blood pressure, glucose, and other factors related to diabetes. The proposed method in this research was evaluated using evaluation metrics such as accuracy, precision, and recall. This model achieved 94.4. % accuracy, 96.5% precision, and 92.7% recall. The results indicate that the proposed model has performed better than other references by increasing the accuracy in diagnosing diabetes. Therefore, by using this model, it will be possible to identify potential diabetic patients more accurately and ultimately prevent them from becoming diabetic.