جستجوی مقالات مرتبط با کلیدواژه « خوشه بندی » در نشریات گروه « برق »

Clustering is an operation in which a set of data samples is categorized according to the degree of similarity. Examples of clustering data are numerical or a mixture of numerical and non-numerical (nominal) data. Finding similarities and measuring distances is one of the challenges of mixed data clustering. In the related works, to detect the degree of similarity and obtain the distance value, only the parameter of the distance value was considered and the cluster was selected based on its value. Clustering in this way, especially for mixed data, has not had very accurate results. In this paper, we have tried to pay attention to the parameter "number of similar features" in calculating the degree of similarity and determining the distance. In assigning each sample to a cluster in cases where the distances are equal or close, the number of common features of the samples will determine the appropriate cluster. That is, we will pay attention to the "number of similar features" in addition to the distance to select the cluster. This idea believes that in cases where the distance of the cluster centers is close to the data object, it is better to choose the cluster center that has more features similar to the data object. Logically and also according to the proposed algorithm, the amount of similarity should be in a larger number of features, not just a few limited features but with high similarity. The parameter of the "number of similar features" has a specific definition and is obtained with a suitable threshold. If the distance value of two features is less than the threshold, those two features are considered as similar features. To calculate the distance in the algorithm, the normalized numerical difference for numerical properties and the Hamming distance for non-numerical properties are used. Determining the initial cluster centers, like many methods, is done randomly, and in subsequent iterations of the algorithm, more appropriate samples are selected as the cluster centers. The algorithm is compared with 5 other algorithms in 5 datasets. In examining the results, three criteria of Accuracy, RI and F-Measure have been used. According to the test results, in the mixed and integer datasets, the algorithm performs at least two percent better than the two algorithms and one percent better than the other algorithm. In another data set, the proposed algorithm had results equal to or close to one percent better accuracy than the superior algorithm. In the last data set, the proposed algorithm was ranked second among 5 algorithms. In general, the proposed algorithm won the top rank in most of the results, and in the rest of the cases, it won the second rank out of the five tested algorithms.

Face recognition poses challenges in accuracy, memory efficiency, and computational complexity. This study proposes a two-stage, three-module approach: Subnetwork modules, Cluster-Finder unit, and Final-Decision module. Unlike random distribution methods, our approach employs clustering for distribution. Each subnetwork, a supervised deep neural network, is trained with cluster-specific data. The Cluster-Finder unit compares test data similarity with each subnetwork’s representative. The Final-Decision module selects the best class. Results indicate superior accuracy, recall, and F1 score compared to competitive methods. The approach is faster and more accurate than non-distribution methods, with comparable speed and higher accuracy than random distribution methods. Experiments on VGGFace2, MS-Celeb-1M, and Glint360K datasets confirm both superior performance and scalability. The proposed method, using KMeans for distribution, outperforms Softmax Dissection and Dynamic Active Class Selection. It simplifies training without additional manipulations, offering efficiency over methodologies like Softmax Dissection and ArcFace parallelization. In conclusion, this study focuses on pre-processing and post-processing without added training complexity. A divide-and-conquer approach addresses accuracy and efficiency challenges. In this study, various sources leading to errors in face recognition systems have been examined. These sources include: imprecise features, overfitting, challenging classes, distribution issues, and decision-making complexities. Various classification scenarios are explored, including non-distributed and models with random and intelligent distributions. Inaccurate features uniformly impact all scenarios, with overfitting posing the greatest challenge in non-distributed scenarios. Challenging classes are better distinguished in intelligent distribution scenarios. Inappropriate distribution has less impact in intelligent scenarios, and decision-making challenges exist in both distributions

Nowadays, high-entropy alloys (HEAs) are a popular domain for researchers which is improved performance by using machine learning (ML). HEAs are formed at least five main elements with close or equal size which is depend on their size and type of elements to extend physical and mechanical features. The ML approach has many applications in various fields. Social network analysis (SNA) is one of the ML tools that is used graph theory. Each graph consists of a number of nodes and edges that each node has its own descriptors. The studies done so far has not used the high-entropy alloys network dataset based on the similarity of content and structural features of each compound. In this paper, a new method is proposed that generalized SNA tools to metallurgical and materials engineering. The proposed method is investigated the HEAs descriptors, in which HEAs descriptors similarity are calculated and the HEAs interaction network is created. The groups have been extracted by Louvain algorithm which each group called cluster. The clusters have alloys with similar properties. The experimental results shown high quality clusters that will be effective in predicting the compounds functionality and discovering new compounds and descriptors. The modularity criterion indicates the quality of the clusters, is about 0.713.

Recommender systems are systems that, over time, learn what product(s) or item(s) each person or customer is (are) likely to like and recommend it (them) to him/her. These systems often operate based on similar behaviors from other (possibly similar) people. Finding similar people is generally a highly time-consuming process due to the large number of users and inaccurate due to the lack of information. For this reason, some methods have resorted to increasing speed. On the other hand, some other methods have added additional information so that they can increase the accuracy of finding similar or neighboring users. Some others have resorted to hybrid methods. Recently, by the use of basic clustering methods, which is based on finding the most similar neighbors with the help of users’ clustering, as well as by using basic content analysis methods and sometimes adding ontology to these methods, researchers have been able to take the advantage of these methods in order to solve some of the above challenges acceptably. In the proposed hybrid recommender system, we have used a two-stage system in which, in the first stage, two models of predictions are made, then in the second stage, by a combining component, the results of the first two parts are combined and the obtained results are given to us as the final results of the system. In the first part, a system based on imputation of missing values fills in the blanks in the scoring matrix. For this end, among the methods of the missing data imputation, we designed a method that was compatible with filling the data set in very sparse conditions, and then generalized it to our own method. In this regard, we have proposed a method based on the grey distance clustering. In the second part, which itself is a hybrid ontology-based recommender system, we first extract the information of each item with the help of a web crawler, then based on a basic article, we produce our own limited ontology, and after that we apply our proposed method. Then, with the help of a proposed method, we improve the ontology structure, thus increasing the accuracy of measuring semantic similarity between the items and users in later stages, and significantly improving the effectiveness of the created recommendations. It should be noted that this ontology is not comprehensive. Finally, we measure the similarity of item-items, user-users, and user-items using an innovative basic ontology similarity measurement method. By the use of this similarity matrix, we cluster users and items, and then store similar users and items as a new feature in the user/item profile for each user/item. This will help us speed up the process of looking for similar users and similar items in the future. In fact, based on this feature, we have increased the speed of the whole work. Since we have set our goal to build a system that makes a balance between the two criteria of accuracy and speed, we use these two criteria to evaluate the proposed system using a real data set. The results of comparing our proposed method with some up-to-date similar methods presented in this field (using the same data set) implies that our method is slower than fast methods, although it is more accurate than them. These results also suggest that the proposed method is faster than accurate methods and its quality is more competitive or even better than them.

The increase of cameras nowadays, and the power of the media in people's lives lead to a staggering amount of video data. It is certain that a method to process this large volume of videos quickly and optimally becomes especially important. With the help of video summarization, this task is achieved and the film is summarized into a series of short but meaningful frames or clips. This study tried to cluster the data by an algorithm (K-Medoids) and then with the help of a convolutional graph attention network, temporal and graph separation is done, then in the next step with the connection rejection method, noises and duplicates are removed, and finally summarization is done by merging the results obtained from two different graphical and temporal steps. The results were analyzed qualitatively and quantitatively on three datasets SumMe, TVSum, and OpenCv. In the qualitative method, an average of 88% accuracy rate in summarization and 31% error rate was achieved, which is one of the highest accuracy rates compared to other methods. In quantitative evaluation, the proposed method has a higher efficiency than the existing methods.

In most applications of wireless sensor networks, it is not possible to charge the nodes' batteries, so the protocols designed for these networks must be as energy-efficient as possible. Clustering is one of the main approaches to designing energy-efficient and scalable protocols for wireless sensor networks. The use of clusters reduces the communication overhead caused by data transmission as well as energy consumption and wave interference between nodes. Despite the importance of clustering in wireless sensor networks, no criteria have yet been proposed to evaluate the quality of clusters derived from clustering algorithms. This paper defines several criteria for evaluating the quality of clusters formed in different clustering protocols. Then, these criteria are combined using fuzzy logic. With the help of the resulting fuzzy criterion, the quality of clusters formed in different clustering algorithms can be better compared. Finally, the correctness and feasibility of this fuzzy evaluation criterion have been verified by simulating three applied protocols and comparing the metrics evaluation results with what is actually happening.

Nowadays, we are witnessing financial markets becoming more competitive, and banks are facing many challenges to attract more deposits from depositors and increase their fee income. Meanwhile, many banks use performance-based incentive plans to encourage their employees to achieve their short-term goals. In the meantime, fairness in the payment of bonuses is one of the important challenges of banks, because not paying attention to this issue can become a factor that destroys the motivation among employees and prevents the bank from achieving its short-term and mid-term goals. This article is trying to tackle the problem of optimizing the coefficients of branch performance evaluation indicators based on their business environment in one of the state banks of Iran. In this article, a two-objective genetic algorithm is proposed to solve the problem. This article is comprised of four main sections. The first section is dedicated to the problem definition which is what is our meaning of optimizing the importance coefficients of branches based on the business environment. The second section is about our proposed solution for the defined problem. In the third section, we are comparing the performance of the proposed two-objective genetic algorithm on the defined problem with the performance of four well-known multi-objective algorithms including NSGAII, SPEAII, PESAII, and MOEA/D. And finally, the set of ZDT problems which is a standard set of multi-objective problems is taken into account for evaluating the general performance of the proposed algorithm comparing four well-known multi-objective algorithms. Our proposed solution for solving the problem of optimizing branch performance coefficients includes two main steps. First, identifying the business environment of the branches and second, optimizing the coefficients with the proposed two-objective genetic algorithm. In the first step, the k-means clustering algorithm is applied to cluster branches with similar business environments. In the second step, to optimize the coefficients, it is necessary to specify the fitness functions. The defined problem is a two-objective problem, the first objective is to minimize the deviation of the real performance of the branches from the expected performance of them, and the second objective is to minimize the deviation of the coefficients from the coefficients determined by the experts. To solve this two-objective problem, a two-objective genetic algorithm is proposed. In this article, two approaches are adopted to compare the proposed solution performance. In the first stage, the results of applying the proposed two-objective genetic algorithm have been compared with the results of applying four well-known multi-objective genetic algorithms on the problem of optimizing the coefficients. The results of this comparison show that the proposed algorithm has outperformed the other compared methods based on the S indicator and run time, and it is also ranked second after the NSGAII algorithm in terms of the HV indicator. Finally, for evaluating the performance of the proposed algorithm with other well-known methods, the set of ZDT problems including ZDT1, ZDT2, ZDT3, ZDT4, and ZDT6 has also been taken into consideration. At this stage, the performance of the proposed algorithm has been compared with the four mentioned algorithms based on four key indicators, including GD, S, H, and run time. The results show, the proposed algorithm has outperformed significantly in terms of run time in all five ZDT problems. In terms of GD indicator, the performance of our proposed algorithm is located in the first or second rank among all considered algorithms. In addition, in terms of S and H indicators in many cases, the proposed algorithm outperformed the other well-known algorithms.

Calculation and applying trust among users has become popular in designing recommender systems in recent years. However, most of the trust-based recommender systems use only one factor for estimating the value of trust. In this paper, a multi-factor approach for estimating trust among users of recommender systems is introduced. In the proposed scheme, first, users of the system are clustered based on their similarities in demographics information and history of ratings. To predict the rating of the active user into a specific item, the value of trust between him and the other users in his cluster is calculated considering the factors i.e. time, location, and context of their rating. To this end, we propose an algorithm based on beta distribution. A novel tree-based measure for computing the semantic similarity between the contexts is utilized. Finally, the rating of the active user is predicted using weighted averaging where trust values are considered as weights. The proposed scheme was performed on three datasets, and the obtained results indicated that it outperforms existing methods in terms of accuracy and other efficiency metrics.

In the Corona crisis, we face a wide range of thoughts, feelings, attitudes, and behaviors on social media. This data contains valuable information for responding to the crisis by the people and administrators. The goal of this study is to identify the characteristics of messages that lead to different emotional polarities. This study aims to investigate the information posted by Twitter, Instagram, and Telegram users and news related to the COVID-19 pandemic in Iran. The data extracted from social networks are focused on the period of January 21, to April 29, 2020, which were shared in Iran and in Persian. It should be noted that the data set and their labels were published by the Cognitive Sciences and Technologies Council (CSTC) in Iran. In this work, the content of each post was pre-processed. Pre-processing was performed by removing stop words, normalizing the words, tokenizing, and stemming. The emotion labels were based on plutchik’s model and included joy, trust, fear, surprise, sadness, anticipation, anger, disgust, stress, and other emotions. In this study, clustering algorithms were used to analyze social media posts. We applied a two-stage clustering method. The proposed clustering algorithm was a combination of self-organized neural network and K-means algorithms. According to our proposed algorithm, the data were clustered through SOM at first, the results of which provided the initial cluster centers for the K-means algorithm. Implementations were built in Python version 3.7 and MATLAB R2015a. Hazm Tools was used for pre-processing data, and clustering was done in MATLAB. The Davies-Bouldin clustering evaluation was applied to find the optimal number of clusters. This measure was calculated for the number of clusters in the range of 2-50 in the two-stage clustering method. The results showed that the optimal number of clusters was ten. Analysis of the results showed that posts related to health and culture with negative polarity led to negative emotions such as fear, hatred, sadness, and anger. Messages about people's emotional and improper functioning have led to feelings of sadness, fear, and stress, and reduced hope in society. The results revealed a strong correlation between anger and disgust. Also, a positive correlation between fear, stress, and sadness was observed. In order to reduce the negative feelings and to create a sense of trust in the authorities, we suggest clarifying about the corona pandemic

Due to the increasing information and the detailed analysis of them, the clustering problems that detect the hidden patterns lie in the data are still of great importance. On the other hand, clustering of high-dimensional data using previous traditional methods has many limitations. In this study, a semi-supervised ensemble clustering method is proposed for a set of high-dimensional medical data. In the proposed method of this study, little information is available as prior knowledge using the information on similarity or dissimilarity (as a number of pairwise constraints). Initially using the transitive property, we generalize the pairwise constraints to all data. Then we divide the feature space into a number of sub-spaces, and to find the optimal clustering solution, the feature space is divided into an unequal number of sub-spaces randomly. A semi-supervised spectral clustering based on the p-Laplacian graph is performed at each sub-space independently. Specifically, to increase the accuracy of spectral clustering, we have used the spectral clustering method based on the p-Laplacian graph. The p-Laplacian graph is a nonlinear generalization of the Laplacian graph. The results of any clustering solutions are compared with the pairwise constraints and according to the level of matching, a degree of confidence is assigned to each clustering solution. Based on these degrees of confidence, an ensemble adjacency matrix is formed, which is the result of considering the results of all clustering solutions for each sub-space. This ensemble adjacency matrix is used in the final spectral clustering algorithm to find the clustering solution of the whole sub-space. Since the sub-spaces are generated randomly with an unequal number of features, clustering results are strongly influenced by different initial values. Therefore, it is necessary to find the optimal sub-space set. To this end, a search algorithm is designed to find the optimal sub-space set. The search process is initialized by forming several sets (we call each set an environment) consisting of several numbers of sub-spaces. An optimal environment is the one that has the best clustering results. The search algorithm utilized three search operators to find the optimal environment. The search operators search all the environments and the consequent sub-spaces both locally and globally. These operators combine two environments and/or replace an environment with a newly generated one. Each search operator tries to find the best possible environment in the entire search space or in a local space. We evaluate the performance of our proposed clustering schema on 20 cancer gene datasets. The normalized mutual information (NMI) criterion and the adjusted rand index (ARI) are used to evaluate the performance evaluation. We first examine the effect of a different number of pairwise constraints. As expected, with increasing the number of pairwise constraints, the efficiency of the proposed method also increases. For example, the NMI value increases from 0.6 to 0.9 on the Khan-2001 dataset, when the number of pairwise constraints increases from 20 to 100. More number of pairwise constraints means more information is available, which helps to improve the performance of the clustering algorithm. Furthermore, we examine the effect of the number of random subspaces. It is observed that increasing the number of random subspaces has a positive effect on clustering performance with respect to the NMI value. In most datasets, when the number of sub-spaces reaches 20, the performance of the proposed method does not change much and is stable. Examining the effect of sampling rate for random subspace generation shows that the proposed method has the best performance in most cancer datasets, such as Armstrong-2002-v3, and Bredel-2005 datasets, when the random subspace generation rate is 0.5, and by deviating the rate from 0.5, the level of satisfaction decreases. Then, the results of the proposed idea are compared with the results of the method proposed in the reference [21] according to ARI and we see that our proposed method has performed better in 12 data sets out of 20 data sets than the method proposed in the reference [21]. Finally, the proposed idea is compared with some metric learning approaches with respect to NMI. We have observed that the proposed method obtained the best results compared to other compared methods on 11 datasets out of 20 datasets. It also achieved the second-best result on 6 out of 20 datasets. For example, the value NMI obtained in the proposed method is 0.1042 more than the reference [21] and it is 0.1846 more than RCA and it is 0.4 more than ITML and also it is 0.468 more than DCA on the Bredel-2005 dataset. Utilizing ensemble clustering methods besides the confidence factor improves the ability of the proposed algorithm to achieve better results. Also, utilizing the transitive operators as well as the selection of random subspaces of unequal sizes play an important role in achieving better performance for the proposed algorithm. Using the p-Laplacian spectral clustering method produces a better, more balanced, and normal volume of clusters compared to the standard spectral clustering. Another effective approach to the performance of the proposed method is to use search operators to find the best subspace, which leads to better results.

Due to the high share of electricity consumption in the country's industries, in recent years, various projects have been implemented, including reducing the amount of load by implementing nationwide blackouts during peak consumption times. Today, data mining is widely used as a process of discovering useful patterns from the database and one of the effective methods for analyzing, modeling and predicting energy consumption. In this study, an integrated clustering-association data mining model has been designed to investigate power consumption behavior to discover and extract the pattern from the power consumption data set of industrial units located in one of the industrial towns of Tehran province. Observations show that during the warm months of the year, the average consumption of high-consumption cluster units, which includes about 34% of the studied industrial units, is about 4.2 times the consumption of low-consumption clusters and about 1.7 times the consumption of medium clusters. By using the proposed model in this study, in addition to identifying high-consumption units and implementing smart and fair policies in forced shutdowns, it is possible to prevent damage caused by forced shutdowns and industrial units can be encouraged to optimize energy consumption. The innovative approach of this model is able to control large volumes of data for planning different areas with the aim of optimizing its energy consumption.

The goal of recommender system is to provide desired items for users. One of the main challenges affecting the performance of recommendation systems is the cold-start problem that is occurred as a result of lack of information about a user/item. In this article, first we will present an approach, uses social streams such as Twitter to create a behavioral profile, then user profiles are clustering with machine learning techniques. Based on this clustering, predictions are made using machine learning techniques such as the Random Forest algorithm (RF) and the Gradient Boosting algorithm (GB). Therefore, the user is not required to provide any kind of data explicitly anymore. As a result of this method, cold start problem will decrease among users' social networks. Because the system uses this data to create user profiles, this will be an input for recommender systems. Numerous experiments have been performed in this field and compared to some new cold start algorithms; very satisfactory results have been obtained. In this paper, we have concluded that the clustering process greatly increases the performance accuracy of the models and reduces the average absolute error, and also the Gradient Boosting algorithm has a better performance than the Random Forest algorithm.

Due to developments of radars in changing parameters, pulse separation methods in ELINT systems have relied on intra-pulse modulation instead of analyzing common pulse description words. Number of radars in this method, may be estimated incorrectly because a modern radar with the ability in changing intra-pulse modulation type may be detected as multiple radars.The purpose of this paper is to detect multi-mode radars with various types of internal modulation in a dense radar environment. The proposed’solution is to add multi-mode radars detection to the existing pulse separation methods. In this method, by extracting the distinguishing features of each radar, similarity criteria of each feature are calculated to examine the similarity between the two pulse streams. Input data are features that  extract of  real data of a ELINT system with traditional parameters and developed parameters related to the pulse shape provided by a simulator. The dataset is then used to train a LVQ neural network to discriminate between different and multi-mode radars. After training the network, in the new operating conditions, a multi-mode radar’s pulse streams is classified as a single radar.The simulation results show a higher accuracy for classification with the similarity criteria of developed features than the criteria extracted of  classical data at different SNRs. Also, the increase in classification accuracy in single-layer perceptron and multi-layer neural network at SNR equal to 0.5 dB has been shown. In similar articles, only multi-mode radars with the ability to change frequency and PRI have been investigated, but with the proposed method, radars with the ability to change several parameters along with the type of intra-pulse modulation can be identyfied.The innovation of the article is the development of distinguishing features and the idea of ​​similarity criteria for the detection of multi-mode radars.

In recent years, Vehicular Ad-hoc Networks (VANET) as an emerging technology have tried to reduce road damage and car accidents through intelligent traffic controlling. In these networks, the rapid movement of vehicles, topology dynamics, and the limitations of network resources engender critical challenges in the routing process. Therefore, providing a stable and reliable routing algorithm is a necessary requirement to maintain the Quality of Service (QoS) parameters of VANETs. In this paper, a new routing algorithm based on the clustering technique is proposed, which is called GCAR. In the proposed algorithm, the appropriate cluster heads are selected based on the genetic algorithm then two vehicles are selected between the neighboring clusters as the gateways and a vehicle chain is formed by these vehicles. Moreover, a combination of genetic algorithm and simulated annealing is applied to identify the suitable clusters. The conducted simulations in MATLAB tool indicate that, respectively, path discovery ratio, the number of clusters, throughput, and packet delivery ratio of the proposed algorithm have been improved by 18.4%, 2.55%, 3.45%, and 14.18% in comparison to the PassCAR approach. Furthermore, evaluation of the convergence, standard deviation, and standard error of the proposed algorithm prove its high convergence speed and stability.