نشریه پژوهش های نظری و کاربردی هوش ماشینی
پیاپی 1 (بهار و تابستان 1402)

Multi-label learning algorithms face many challenges due to the high volume and dimensions of multi-label data and the existence of noise. Feature selection methods are an effective technique for addressing these challenges. This paper presents a feature selection method based on an ensemble approach for multi-label data. In this approach, three different decision matrices based on various feature evaluation criteria, taking into account the relevancy of features with class labels and their redundancy relative to each other, are effective in the feature selection process. These three decision matrices are finally combined based on an ensemble approach using the concept of fuzzy integral to evaluate the features according to the aggregate value. Comparisons have been made with several similar algorithms to illustrate the performance of the proposed method.

In the use of face recognition systems, various fraud, such as the use of a mask and a photo of a genuine person, are two major problems that limit their applications. Studies have shown a number of methods for detecting fraud in face recognition, which are sometimes intrusive, enforcing the person to make a move in order to distinguish the real face from the fake one. The use of intrusive methods often leads to user dissatisfaction. In this article, we present a non-intrusive method using features such as light reflection or the presence of periodic noise to distinguish real images from the fake one. In this method, the edges and texture of the image are highlighted by a local binary pattern to better detect fraud. Then, by extracting the image feature using a deep learning technique with three layers of convolution, it will be able to distinguish between real and fake face images. This method is resistant to covering the eyes and face. In order to evaluate the proposed method, the CASIA dataset was used in this research. The results show 98% accuracy of the proposed method on this dataset. Among the existing methods, we see an increase in accuracy.

Part of speech tagging is an important issue in natural language processing and is the base of many other major subjects in this field. In this article, a new method have been introduced for part of speech tagging using deep neural networks. The purpose of this method is solving problems that common other methods are facing with which are extract deep features from texts and classifying these features. The proposed method is based on that we can find deep features and product optimal output by using small deep neural networks. This method was implemented using Tensorflow's specialized libraries and Keras API in python and was evaluated on coNLL2000 standard dataset. The experimental results show that the proposed method is capable to extract high level features from natural language's words and is able to achieve considerable accuracy for repetitive tags. In addition, this method is able to be used in variant environments and on different devices.

Schizophrenia is a disorder of the brain in which people misinterpret reality. This mental disorder is characterized by behavioral symptoms such as hallucinations and speech disorders. Electroencephalogram (EEG) signal indicates brain disorders and is widely used to study brain diseases. The aim of this article is to automatically detect schizophrenia based on the EEG signal. A common method in research is the manual extraction of features from the EEG signal. Because deep learning algorithms have the ability to automatically extract important features and classify them, in this study, in order to extract more useful features, the EEG signal was applied to an eleven-layer convolutional deep recursive neural network. The EEG signals collected at the Warsaw Institute from 14 healthy individuals and 14 patients with schizophrenia are studied here. The mean values of the accuracy evaluation criteria of the model including Accuracy, Sensitivity, Specificity and PPV for the proposed model were 98.79%, 98.73%, 98.86% and 99.06%, respectively, which improved the performance of the proposed model for classifying schizophrenic patients and approves healthy people compared to previous models. The proposed model can be used as a diagnostic tool to help physicians diagnose the early stages of schizophrenia.

One of the problems with high-dimensional data is choosing the best features, because all the features of the data to find the knowledge that the data lies are not important and vital. For this reason, reducing the size of the data is one of the important issues. Hence in This research has tried a new method using sine cosine algorithm with multiple optimization approach in the Feature selection field. In fact, the innovation of this research is in providing a way to obtain the whole set of appropriate features, which for the first-time sine cosine algorithm has been improved.The proposed method is presented in the wrapper feature selection model and has two steps, which include the feature selection step using the multimodal sine cosine algorithm and the classification step of possible solutions obtained from sine cosine algorithm by the extended nearest neighbor classification method.The proposed method was tested on data sets from uci with different dimensions. The results of the proposed method along with the results of other methods including multimodal optimization and single optimizations are compared and it is observed that the proposed method compared to the single optimization methods, has higher efficiency and compared to multimodal optimization methods, it had better result with a slight difference.In general, the proposed method has been able to reduce the number of features by more than 5% compared to other methods and the average accuracy of the classification compared to the best results of other methods has improved by an average of 2%.

Visual object tracking in arbitrary environments with arbitrary objects has gained considerable importance in recent years. A very significant feature, which makes a tracker useful, is real time tracking without needing GPU and pre-train algorithms. In the recent decade, the trackers, which function on the basis of discriminative correlation filters, have promised positive results in terms of both speed and accuracy. Although, in most of such methods, the estimation of the position of the object in each frame is computed based on transformation and pyramid scales, in Large Displacement Estimation of Similarity transformation algorithm, translation, scale and rotation are estimated in each frame. In this paper, the Histogram of Oriented Gradient is considered as feature extraction. Here, we adopt two different approaches. The first approach uses scaled images as a feature matrix by applying minimum variance quantization. The second approach, uses a combination of opposite color local binary patterns and Speeded-Up Robust Features. By using these two methods, we are able to extract helpful and fast features, and therefore improve the results of tracking against challenging attributes. The OTB-2015 dataset is utilized for evaluating tracker. The results show precision of trackers improve 3%. Additionally, the first tracker increase the result about 7% against low resolution and the second one can be helpful about 4% in facing rotation challenge.

In some real-world applications, there is high-dimensional data which has led to many computational challenges. Feature selection is an effective technique for data dimensionality reduction, which simplifies the model and improves its performance by selecting the appropriate subset of features. In many of these applications, labeling of data is costly and time consuming, leaving little labeled data available and large amounts of unlabeled data available. In such applications, semi-supervised feature selection methods perform the feature selection process using the information of labeled data, and the distribution and geometric structure of labeled and unlabeled data. In most semi-supervised feature selection methods, a neighborhood graph is created and the importance of features is evaluated via their ability to maintain the geometric structure of the graph. In classical graph-based semi-supervised feature selection methods, the features are evaluated one by one and the correlation between features is not considered in feature selection process. To overcome this problem, sparse feature selection methods have been presented which consider the correlation between features, and calculate the optimal sparse transformation matrix for feature selection. In this paper, we investigate the semi-supervised learning methods, and review the graph-based semi- supervised sparse feature selection methods which select the appropriate features using the graph created by the labeled and unlabeled data, and the sparse regularization term. These methods solve the problem of classical semi-supervised methods by considering the correlation between features, create a neighborhood graph using the labeled and unlabeled data, calculate the graph Laplacian matrix, and compute the optimal sparse transformation matrix for feature selection.

Clustering or community detection is a powerfrul tool for ‎analayzing complex networks which is widely used for ‎modeling complex systems. Modularity is a ‎comprehensive criterion for evaluating the quality of ‎clusters (or communities). However, it has some ‎limitations and challenges such as being a NP-hard ‎problem and not using prior information. So, Modularity-‎based community detection cannot be extended as a ‎semi-supervised community detection method. On the ‎other hand, one of the most common semi-supervised ‎methods which can use prior knowledge for clustering is ‎community detection based on non negative matrix ‎factorization (NMF). But, this method is not able to ‎consider the features of the networks. Therefore, in this ‎paper to overcome the mentioned limitations and ‎challenges and by presenting a new proof, a structure ‎similar to community detection based on NMF is ‎presented for modularity-based community detection ‎which can employ prior knowledge and iterative ‎solution. Therefore, a novel semi-supervised community ‎detection based on modularity (SSNMF-Q) criteria is ‎developed by utilizing prior information and iterative ‎solution instead of solving a NP-hard problem. To ‎evaluate SSNMF-Q, five real world networks are used ‎and it is shown that the SSNMF-Q had better ‎performance compared to other semi-supervised ‎community detection methods based on NMF.‎

Trees are one of the most basic classes in graphs. They not only play a key role in the theory of graphs and combinatorics, but also in many other fields of mathematics, as well as in other sciences such as biology, chemistry, and computer science. In this article, we review the applications of trees in chemistry, biology, and computers.

Today, due to the large volume of opinions published by people in cyberspace, sentiment analysis plays a key role in extracting information. One of the new techniques based on studies has been done to determine the exact polarity of the sentence in sentiment analysis is deep learning algorithms. In this research, two deep learning algorithms, namely RNN and LSTM, has been used to determine sentence polarity in order to achieve more accurate results. Moreover, in the proposed technique, pre-trained word embedding algorithm, namely Wordtovec, was used to determine the semantic relationships between words to increase the accuracy of the proposed method. The proposed method was evaluated on two data sets; airline-tweet and IMDB. The evaluation results show that on the airline-tweet dataset, the proposed method has an accuracy of 0.78 and accuracy of 0.84 on the IMDB data set.

Facial skin segmentation plays an important role in applications such as identification, facial expression analysis, facial animation, and skin disease analysis. Clustering is one of the most common methods for image segmentation. In this paper, a new hybrid method based on Quantum Particle Swarm Optimization and Grey Wolf Optimization is presented to optimize the performance of the K-Means clustering. By Combination of two algorithms, the exploitation performance of the QPSO algorithm is improved by the exploration capability of the GWO algorithm. To measure the similarity, four distance criteria including Euclidean, Minkowski,  Mahalanobis, and City-Block distances have been used to optimize the K-Means algorithm. The proposed method has a better performance in segmentation and convergence speed compared to other meta-heuristic algorithms such as Genetic Algorithm, GWO, PSO, QPSO, Bat Optimization, Crow Search Algorithm. The experimental results show that Minkowski distance has a better performance in calculating similarity and optimization of K-Means algorithm. Based on the obtained results, the proposed method ensures the achievement of the optimal solution and prevents the problem from falling to a local minimum.

Identifying and eliminating throughput bottlenecks is a key tool to increase throughput and productivity in production systems. However, due to the complexity and dynamics of factory, eliminating throughput bottlenecks is considered a major challenge. Researchers have tried to develop tools to help identify and eliminate these bottlenecks. Historically, research efforts have focused on developing modeling approaches to identify bottlenecks in manufacturing systems. However, with the advent of industrial digitization and artificial intelligence, researchers have explored various ways in which artificial intelligence can be used to eliminate bottlenecks. In this research, the role of artificial intelligence in identifying and eliminating bottlenecks is stated and the efforts made in the field of operational throughput bottlenecks are classified into four categories: (1) identify, (2) diagnose, (3) predict and (4) prescribe. Also, practical recommendations and future research topics have been provided, which can help to improve the practi c a l a n d t h e o r e t i c a l u s e o f a r t i f i c i a l i n t e l l i g e n c e i n i n d u s t r i e s .