فصلنامه پردازش علائم و داده ها
سال بیستم شماره 3 (پیاپی 57، پاییز 1402)

Predicting and estimating the time it takes for an event of interest to occur base on available information is special assistance in how to deal with the event and handle it or provide solution to prevent the occurrence of the event. In medicine, valuable information about evaluating the types of treatments and prognosis and providing solution to handle event can be gained by predicting the time that an event occurrence according to information recorded from patients. Many statistical solutions have been proposed for predicting the time that an event occurrence and the most professional method is Survival Analysis. The purpose of Survival Analysis is to predict the time that an event occurrence a model effective parameters in estimating the time, which can be control or eliminating problematic factors. Due to the importance and prevalence of breast cancer as the second leading cause of death among cancer patients in the world, access to models that can accurately predict the survival of breast cancer patients is very important. The present study is an analytical study. The data used in this study are taken from The Molecular Taxonomy Data of the International Federation of Breast Cancer (METABRIC) database, which is related to which is related to the molecular classification of breast cancer patients. The total number of patients studied was 1981. Of these, 888 patients were in care until the time of death and the rest did not continue the study during the study. In this database, 21 clinical features of patients have been considered, which includes a total of 6 quantitative features and 15 qualitative features. To predict survival, a deep neural network model called the optimized DeepHit is used. The optimized model has achieved the criterion of c_index = 0.73, which is a criterion for measuring the capability of survival analysis models. Comparisons with previous models based on real and synthetic datasets show that the optimized DeepHit has achieved great performance and statistically significant improvements over previous advanced methods.

Subject: 

Today, with the advancement of technologies to assist blind and visually impaired people, navigation systems are of great importance. As a result of emerging technologies in telecommunication and smartphones, these people can be helped. Identifying pedestrian and traffic lights is important to help pedestrians with visual impairments cross the intersection safely and securely. Background- researchers have studied the detection and identification of traffic lights in the assistive system or blind assist device. These researches can be divided into three main types: based on pattern matching, based on circular shape extraction, and based on color distribution.

In this research, an architecture based on mobile cloud computing is proposed, which can help blind pedestrians in crossing intersections. The architecture consists of three tiers: mobile phone, cloud, and supervision. The most important component is located on the mobile phone. It recognizes the color of pedestrian light by using image processing techniques. Spatial information (time and location) of the blind person is collected and held in a cloud storage database so that acquaintances can monitor him if needed. In order to detect the status of pedestrian lights, pictures of crossing streets with cameras will be captured. Using the features of color and morphology operations, the color of pedestrian lights is recognized and reported to the blind person. To this end, morphological operations are performed to eliminate small elements in the background and to restore the original size of the traffic light sign. Therefore, the operations of dilation, filling, and erosion are used.

We gathered a dataset including 280 photos of pedestrian lights (170 photos at day, 110 photos at night) in different illumination conditions (early day, noon, early night, night) and weather (sunny, cloudy, rainy). Matlab software and notebook system with Intel (R) Core (TM) i5 CPU and AMD Mobility Radeon HD 5100 graphics card were used to implement pedestrian traffic-light status detection. The scenario-based method is used to evaluate the system architecture and show that the proposed system can satisfy the investigated scenario. At last, the implementation results on taken images show excellent performance in detecting pedestrian lights with approximately 100% accuracy for the day and night.

Coronavirus (Covid-19) is a new infectious disease with a very high rate of infection and mortality. Therefore, its early detection has become one of the vital measures of human society. This virus is commonly tested using sputum or blood samples, and the result is usually announced within hours or even days after the test. However, even an hour delay in announcing the test results can lead to many more people being infected. Another way to diagnose the virus is to take a chest X-ray (chest radiography images), which is much faster and cheaper than other tests. However, the rate of human diagnostic error from these images is high, and more importantly, it is very difficult to distinguish Covid-19 infection from other infections such as pneumonia. In this paper, an intelligent system, termed CoviX-Net, based on deep machine learning techniques, is presented to diagnose and differentiate Covid-19 disease and various types of pneumonia (bacterial or viral) using chest radiographs. The CoviX-Net learning model is based on the Xception architecture, the accuracy of which is improved by the use of transfer learning and data augmentation techniques. To provide adequate training data, a comprehensive database is created by integrating two different sources of chest X-ray images. The evaluations performed on the test images show that the accuracy of CoviX-Net diagnosis in the three-class mode (Covid-19, pneumonia, and normal lung) is %99.25, and in the four-class mode (Covid-19, bacterial pneumonia, viral pneumonia, and normal lung) is %95. Compared to other similar deep learning-based classification methods, the accuracy is improved by 5%, and compared to the transfer learning-based parallel deep learning method, with a complex structure, the accuracy is improved by about half a percent. These promising results demonstrate the superiority of CoviX-Net against the competitors, and suggest that CoviX-Net can be a useful tool to assist clinicians and radiologists in diagnosing patients with Covid-19. All the implementation source codes and collected dataset are made publicly available.

Nowadays, social networks play an important role in people's daily lives. A social network is a kind of social structure that consists of several nodes that can be individuals or organizations. Most importantly, these nodes are connected by one or more specific types of dependencies, such as friendships or work relationships. Understanding the structure and constituent groups of these networks can give us useful information about the state of society and individuals. In this article, a new solution to solve the problem of social structure detection is provided. Social structure means communities or associations in social networks. An important issue of this context is network graph construction based on objects as nodes and edges as transactions between these objects. Community detection is based on these graphs. The appropriate solution is to identify and create clusters of nodes that have strong connections with each other and at the same time have weaker connections between nodes of different clusters. Optimization algorithms can be used to construct and detect these connections. Harmonic search is one of the efficient optimization algorithms in this context. However, in the field of identifying communication structures and communities, so far, no research work has been done using the harmonic search algorithm. In this paper, a new method is proposed to construct network clusters based on network graphs. It can identify effective communications based on different criteria. In order to propose this method, first a new version of harmonic search algorithm is designed for discrete environments while the original version of which is for continuous environments. Then, according the problem, which is to discover appropriate structures in the social network graph, a new method is devised to solve it. This method tries to provide a suitable discrete version by relying on different operators to be applicable to solve the desired problem. In order to evaluate the proposed method, various experiments were carried out on several different networks. These networks have been used as benchmark in previous research work. For evaluation and comparison, two artificial networks and two real networks are considered. The evaluation results of the proposed method on these networks are presented based on different criteria and compared with four previous algorithms that are known in this field. Comparison results show that the proposed algorithm is relatively superior to other algorithms or at least produces similar results. The most important reason that can justify the relative performance superiority of the proposed algorithm or at least its competitive results is the better search capability of the problem search space. This leads to the discovery of more promising points and the production of better solutions.

Brain-computer interface systems based on classification of the motor imageries (MI) using multi-channel EEG signal play a major role in the control of artificial limbs and machines by people with disabilities. One of the main problems in classifying these signals to recognize different MI tasks is the large number of channels. The large number of channels causes a lot of cost and hassle during the measurement process, increasing computational load of the preprocessing, feature extraction, and classification, difficulty of interpretation of results, and over-fitting of the classifier due to the limited number and noisy training samples. Since not all measured channels for classifying a particular MI task have useful information, it would be beneficial to select the optimal channels for classifying desired MI tasks. Channel selection methods are categorized into wrapper, filtered, hybrid, and embedded categories. In this paper, a filtering method is used due to less computational cost and the independence of the classifier. The used criterion is very important in filtering methods. Criteria based on first- and second-order data moments are less efficient for non-Gaussian classes. The proposed method uses mutual information between candidate channels and class label as a comprehensive criterion and sequential forward selection search strategy. One of the problems in using this criterion is the accurate estimation of mutual information in the high dimensional spaces. The kpn entropy estimator is used to accurately estimate the mutual information in high dimensional space with limited number of training samples. The power of 2 Hz non overlapping sub-bands in the 8-30 Hz band and in 250 milliseconds non overlapping intervals in half to two and a half seconds after the onset of MI are extracted as features for each channel. The extracted features are reduced to 10 for each channel by combining the unsupervised L1-PCA and supervised NWFE dimensionality reduction methods. The reported results show the ability of the proposed method to select effective channels for classifying left and right hand and feet MI tasks. The overall accuracy of the SVM classifier on test samples for two subjects labeled aa and al from the BCI III competition dataset is 94.87% and 96.51%, respectively, while the number of channels is reduced from 118 to 7 channels.

Key employee's turnover is one of the most important concerns of Human Resource Managers (HRM); Because the organization by losing its valuable staff, suffers from the loss of skills and experience gained over the years, so predicting employee turnover helps HRMs to hire and retain permanent employees. One of the effective tools in this regard is the use of different data mining methods. Many researchers have done research in this field. This study reviewes recently published articles based on machine learning models, using Kaggle Human Resource (HR) databases [1-5] to compare them with this proposed models. In the article [9], the authors have selected 11 of the most important features by collecting common features from previous articles and filtering them using feature review and selection algorithms. After converting non-numerical variables to numerical and normalizing the data in the range [0,1], those attrition prediction approach is based on machine, deep and ensemble learning models and is experimented on a large-sized and a medium-sized simulated HR datasets and then a real small-sized dataset from a total of 450 responses. Those approach achieves higher Accuracy (0.96, 0.98 and 0.99 respectively) for the three datasets when compared previous solutions. In 2021, authors examined the relationship between features using Pearson correlation coefficient and selected 11 features with the highest correlation coefficient. Then used from six different machine learning algorithms including Random Forest (RF), Logistic Regression (LR), …, to predict employee turnover. The highest accuracy they obtained was 0.85 for RF [3]. In the article[1], the authors used two IBM datasets and a database containing HR information from a regional bank in the USA to predict employees turnover. After cleaning and preprocessing the data, the performance of 10 different machine learning algorithms such as Decision Tree (DT), RF, LR, Neural Network, …, was evaluated using ROC criteria on 10 small, medium, and large subsets of randomly selected, unassigned primary datasets. The average accuracy of algorithms is 0.83 in small datasets, 0.81 in medium datasets and 0.86 in large datasets. The authors of the paper [4] used three main experiments on IBM Watson simulated datasets to predict employees turnover. The first experiment involved training the original class-imbalanced dataset with the following machine learning models: support vector machine with several kernel functions, random forest and K-nearest neighbour (KNN). The second experiment focused on using adaptive synthetic (ADASYN) approach to overcome class imbalance, then retraining on the new dataset using the abovementioned machine learning models. As a result, training an ADASYN-balanced dataset with KNN (K = 3) achieved the highest performance, with 0.93 F1-score. this turnover prediction approach is based on tree-based ensemble learning models and is experimented on a large-sized standard simulated HR dataset (hr_data), including 15,000 samples with 10 features and a medium-sized (IBM) including 1470 samples with 34 features. The employees turnover rate in the IBM is 16.1% and in the hr_data is 23.8%, so datasets are unbalanced. To balance the data, the random-under-sampling technique and its combination of random-over-sampling with a ratio of 0.5965 for the IBM and 0.6558 for the hr_data has been used. In the preprocessing stage, Features with zero variance and samples containing the missing value were also removed. Then categorical (non-numeric) values ​​were converted to binary fields and then All features were scaled using data normalization in [0,1]. In order to reduce the feature dimensions in the IBM dataset, we used the "Non-negative Matrix Factorization" (NMF) technique (n_components=17, max_iter=500) and For initialization, non-negative singular value analysis method with zeros filled with X value has been used. After reviewing and cleaning the data, in the processing stage, six different classification algorithms, including KNN (k=1), RF (number of trees= 1500), DT, ExtraTreesClassifier (number of trees= 1000) and Support Vector Classifier were training on 70% of data. The optimal value of the hyperparameters for the algorithms, was set using RandomizedSearchCV and GridSearchCV techniques. In order to investigate the effect of balancing and Dimensionality Reduction on the performance of models, experiments were performed in 3 stages (befor balancing, after balancing befor Dimensionality Reduction, after balancing and Dimensionality Reduction) on 30% of the remaining data. The results shown in Table (2-4) indicate that this proposed model, which uses tree-based optimized ensemble learning algorithms with data balancing and NMF dimensionality reduction method, increases the f1score of turnover prediction. In the hr_data dataset, the best f1score for the RandomForest algorithm was 99.52% and for the IBM HR dataset, the best f1score for the ExtraTreesClassifier algorithm was 95.82%, which is higher than previous research. Table 5 compares the results of previous research with this research. Since, the prediction of employee attrition will not be enough without finding the characteristics that affect it, therefore, after building models and evaluating their performance, using a combined feature selection method by averaging the results of the single-variable feature selection method called "SelectKBest", and A wrapper feature selection method called "Recursive feature elimination" (RFE) with four learning algorithms RF, DT, ExtraTreesClassifier and AdaBoost, the most effective features were selected. SelectKBest combines the chi2 univariate statistical test with the selection of K features based on the statistical result between the features and the target variable. Also, in the RFE method, machine learning algorithms are used to remove the least important features after recursive training, so that finally the number of features reaches the set number (17 features in this article). The performance results of the models based on the selected features are shown in Table 6. The most effective characteristics are "age", "daily rate", "over  time", "NumCompaniesWorked" and, "monthly income" .

The development of information technology and its use in the health system has taken many measures to protect and promote human health, however, the world still faces long-term threats and recurrence of infectious diseases. Understanding the dynamics of infectious diseases is important in controlling the disease because the network and the mode of impact of infectious diseases are very complex. The management of infectious diseases can also be considered as a complex social system due to the fact that has many complexities (such as dimensions, parameters, interactions, behaviors and rules), for this reason, the approach of the present study is a multifaceted understanding of the spread of infectious diseases. To design the present model, an intelligent system with a combination of mathematical, machine learning and epidemiological dimensions is proposed. The disease studied in this study, due to its importance and prevalence, is Covid 19. In this study, with the approach of complex systems and using the Internet of Things and machine learning methods, an algorithm was presented that uses environmental and individual variables to predict the probability of disease in an individual. Therefore, this research can improve the prevention of infectious diseases by filling some of the gaps in 3 sections: 1- Re-emergence of infectious diseases and the potential of IoT and AI, 2- Speed of dissemination and importance of real-time tracking, and 3- Budget and cost. The evaluation of the algorithm in this study was determined by two criteria of sensitivity and specificity. The results of the proposed algorithm for predicting Covid 19 disease showed an accuracy of more than 98%. Sensitivity above 98% was also obtained. Which is very important for the diagnosis of Covid disease 19 and shows the low number of false negatives in the test results. Therefore, the proposed model, combined with the Internet of Things and machine learning, can cause early diagnosis and prevent the spread of the Covid-19 disease with high specificity and sensitivity.

Measuring similarity between two text snippets is one of the essential tasks in many NLP problems and it has been still one of the most challenging tasks in the field. Various methods have been proposed to measure text similarity. This survey reviews more than 150 of the related papers, introduces a comprehensive taxonomy with three main categories, and discusses the advantages and disadvantages of these methods. The first category is lexical methods that only focus on text pair’s surface similarity. These methods consider the text as a sequence of characters, tokens, or a mixture of these two. Some recent studies use deep learning techniques for detecting lexical similarity in alias detection task. The second category is semantic methods that take into consideration the meaning of the words based on some pre-prepared knowledge-bases like Wordnet or using Corpus-based methods. Some recent studies use modern deep learning techniques like transformers and Siamese networks to create document embedding that outperform other methods. The final category is hybrid methods that take advantage of all other methods even syntactic parsing in some cases. Note that high-quality syntactic parsers are not present for many languages and that using them has some side-effects on performance and speed.

In recent years, convolutional neural networks (CNNs) have been increasingly used in various applications of machine vision. CNNs simulate the function of the brain's visual cortex and have a powerful structure for analyzing visual images. However, the diversity of digital images, their content, and their features necessitate that CNN networks are specially designed, and their parameters are carefully adjusted to achieve higher efficiency in any classification problem. In this regard, in many previous studies, researchers have attempted to increase the efficiency of the CNNs by setting their adjustable parameters more accurately.

New method In this study, we presented a novel initializing method for the kernels of the first convolutional layer of the CNN networks. We designed a filter bank with specialized kernels and used them in the first convolution layer of the proposed models. These kernels, compared to the random kernels in traditional CNNs, extract more effective features from the input images without increasing the computational cost of the network, and improve the classification accuracy by covering all the important characteristics.

The dataset used in this paper was the MNIST database of handwritten digits. We examined the performance of CNN networks when three different types of kernels were used in their first convolution layer. The first group of kernels had constant coefficients; the second group had random coefficients, and finally, the kernels of the third group were specially designed to extract a wide range of image features. Our experiments on a single-layer CNN network with three types of kernels (constant numbers, random numbers, and filter-bank) showed the average classification accuracy of MNIST images in 50 times of network training to be 74.94%, 86.47%, and 91.89%, respectively, and for a three-layer CNN network, 88.82%, 96.16%, and 99.14%, respectively. Comparison with existing methods Compared to the kernels with randomized coefficients, the use of specialized kernels in the first convolution layer of the CNN networks has several important advantages: 1) They can be designed to extract all important features of the input images, 2) They can be designed more effectively based on the problem in hand, 3) They cause the training to start from a more appropriate point, and in this way, the speed of training and the classification accuracy of the network increase.

This study provides a novel method for initializing kernels in convolution layers of CNN networks to enhance their performance in image classification works. Our results show that compared to random kernels, the kernels used in the proposed models extract more effective features from the images at different frequencies and increase the classification accuracy by starting the training algorithm from a more appropriate point, without increasing the computational cost. Therefore, it can be concluded that the initial coefficients of the convolution layer kernels are effective on the classification accuracy of CNN networks, and by using more effective kernels in the convolution layers, these networks can be made specific to the problem and, in this way, increase the efficiency of the network.

Steganography is the art of transferring information through secret communication. The essential aim of steganography is to minimize the distortion caused by embedding the secret message; so that the image containing the message (stego) cannot be distinguished from the original image (cover), and the existence of the hidden message cannot be detected. The distortion in content-adaptive steganography depends on the local structure of the image. The embedding changes into the areas with rich textures are less detectable than smooth areas, so the textured areas have a higher modification priority. In this regard, three main steganography approaches are proposed: model-based, cost-based, and adversarial. The model-based approach considers a statistical model for the cover image and tries to preserve this model during the embedding process. The cost-based one focuses on minimizing the distortion obtained from the sum of the heuristic costs of modified pixels. The adversarial approach uses the competition between steganography and steganalysis to improve the embedding performance. In the first section of this paper, the concept of steganography and its history is expressed. Digital steganography including three types of cover synthesis, selection, and modification is introduced in the second section. The focus of this paper is on steganography based on the cover modification. The goal is to estimate the best probability distribution of modifications, and embedding the message in the estimated places is left to existing coding algorithms. In the third section, the problem of estimating the probability distribution is formulated as an optimization problem with the aim of distortion minimization. The distortion-based methods compute the probability distribution of embedding changes using a pre-defined distortion function. In the additive distortion function, the embedding changes are assumed to be independent. Thus, the distortion function cannot capture interactions between changes caused by embedding, and it leads the performance to suboptimality. In this regard, the non-additive distortion functions are presented that consider the dependencies among the modification of adjacent pixels. The distortion-based methods include two model-based and cost-based approaches are introduced in the fourth section. Then, their most significant methods are reviewed in the fifth section. Considering the competitive nature of steganography and steganalysis, a new steganography approach is presented in the sixth section that takes advantage of adversarial learning to improve secrecy. Adversarial learning includes two strategies: Generative adversarial networks (GANs) and adversarial attacks. In the concept of steganography, the GAN-based strategy tries to train the steganographic network against a steganalysis network. This is an iterative and dynamic game between steganographic and steganalysis networks to reach the Nash equilibrium. Another strategy attempts to simulate an adversarial attack and generate stego images that deceive the steganalysis network. The adversarial-based steganography methods are reviewed in the seventh section. In the eighth section, different methods are compared from various points of view. The results of this study show that some techniques, such as smoothing the embedding changes, considering the interactions between the changes, using side-informed information, and exploring adversarial networks, can help to estimate the proper embedding probability map and improve performance and security. In the ninth section, suggestions are stated that can be considered for future research. Finally, the conclusion is expressed in the tenth section.

With the expansion of the Internet and wireless technologies, video-related applications have been expanded much more than ever. But since network bandwidth is limited, better high-gain video compression methods remain essential. Recently, the new video compression standard called "High Efficiency Video Coding" (HEVC) has been introduced, which doubles the compression compared to the previous standard called "Advanced Video Coding", while preserving video quality at the same time. In HEVC standard, 35 intra prediction modes have been defined to improve the efficiency of encoding algorithm. The best intra prediction mode is chosen using a process called "Rate Distortion Optimization". The purpose of this study is to propose a method to reduce the number of intra prediction modes in RDO process, since it is a very time consuming process. A new method based on Sobel operator is proposed for fast determination of intra prediction mode as well as coding unit size in HEVC. Although Sobel operator is an essential part of our algorithm, other features were extracted from coding control part of HEVC algorithm to excel the speed of coding process, since the main purpose of fast coding algorithm should be to reduce complexity while producing a bitstream, which preserves the quality of decoded video. These features are parent CU mode, modes with lower cost functions and modes that are determined to be most probable via pervious researches. Parent CU mode is intra angle of coding unit that contains the current CU and is available at the time of calculating the best mode for current CU. Modes with lower cost functions are modes that have the least cost function value that are calculated by HM via a fast algorithm and modes that are determined to be most probable via pervious researches are DC and horizontal and vertical modes. Coding time compared to HEVC reference software (HM) using the proposed fast method is reduced by an average of 56% and by using the proposed precise method is reduced by an average of 29%. In proposed precise method, final video quality degradation is negligible.

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.