فصلنامه پردازش علائم و داده ها
سال بیستم شماره 4 (پیاپی 58، زمستان 1402)

Trust management systems are used in interactive environments, where an agent needs to make a decision about using a service. Due to the preponderance of these systems, malicious entities have strong incentives to influence trust management systems and divert their decisions. In spite of approaches presented in previous trust models to mitigate the malicious activities, many of them could not cope with the problem efficiently. For example, tackling the variable behavior of agents is a common failure point for many trust models. Moreover, no rigid, flexible and adaptive general approach has been presented and the problem somehow remains.In this paper, a novel approach has been presented to prevent malicious actions and recognize anomalies using an entropy-based trust management system with capability of recognizing intrinsic characteristics of the actions, whether the action is malicious or not. the trust calculation is carried out based on an entropy structure derived from information ethics theory. Using this method, malicious actions disrupting the trust management system are filtered thus increasing the trust calculation accuracy. The experimental results demonstrate that the performance of the proposed system is promising in terms of trust calculation accuracy as well as detection of malicious behavior. specifically, the proposed model achieves a 10 percent advantage to well-known trust models in terms of speed of adaption to environmental changes and varying agent behaviors.

Continual learning from data streams is a pivotal aspect of machine learning, requiring the development of algorithms capable of adapting to incoming data. However, the ongoing evolution of data streams presents a formidable challenge as previously acquired knowledge may become outdated. This challenge, known as concept drift, demands timely detection for the effective adaptation of learning models. While various drift detectors have been proposed, they often assume a relatively balanced class distribution. In scenarios with imbalanced data streams, these detectors may exhibit bias toward majority classes, overlooking shifts in minority classes. Moreover, the imbalance among classes can change over time, with roles shifting between majority and minority classes, especially when relationships among classes become complex due to overlapping regions. In this paper, a novel classification method is introduced for imbalanced streaming data affected by concept drift. The proposed method continuously monitors arriving streams to detect and adapt to both imbalances and concept drift. Upon receiving a new block of data, the proposed method employs the k-means clustering approach to identify non-dense regions and performs oversampling for minority classes. Cluster centers are selected using the belief function to address overlapping issues between majority and minority classes. Utilizing a chaotic approach, the new sample is added based on its neighborhood and the size of that neighborhood. Subsequently, concept drift detection is conducted using three pre-defined thresholds that cover time intervals and classification errors. Finally, the label prediction process is done by ensemble learning and weighted majority voting. Experiments conducted on benchmark datasets from the UCI database evaluate the performance of the proposed method using Leave-One-Out (LOO) validation and comparisons with state-of-the-art methods. The results demonstrate the superiority of the proposed method across various evaluation criteria, highlighting its effectiveness in addressing imbalanced streaming data with concept drift.

In this research, a new approach for image segmentation based on genetic algorithm and random forest is presented. Image segmentation can be done using supervised learning. In this learning, there are a number of images from data set with their labels. In image segmentation, different parts of the image separate from each other. In this process, all the pixels in the image are given a label, so that the pixels with the same label have common characteristics with each other. To provide a model that can perform image segmentation, it is necessary to extract features from input images and perform segmentation using a suitable classifier and these features. Image feature extraction is done using image filters. In this research, a hybrid combination of 4 Gabor filter banks and Sobel, Prewitt, Canny edge, Scharr, Gaussian, median, and Roberts filters are used for effective feature extraction. One of the most important of these filters, which also has a degree of freedom, is the Gabor filter. This filter has a number of   hyperparameters that change the efficiency of the classifier by changing these hyperparameters. In this research, an attempt has been made to adjust these hyperparameters using genetic algorithm. The fitness function proposed in this research is f1-score. random forest classifier is utilized for image segmentation and classification. The results of the experiments show that the hyperparameters found by the genetic algorithm have been able to perform a satisfactory segmentation on data set.

In today’s competitive world, companies need to analyze, identify and predict the behaviorof their customers and respond to their demands earlier than their competitors. Moreover, in manyindustries such as mobile telecommunications, the cost of maintaining existingcustomers (customer retention)is much lower than the cost of attracting a new customer. Therefore,the problem of identifying customers who are going to leave the company, so-called Customer Churn Prediction (CCP),and preventing themby offering Incentivesis essential in these industries. In this direction, researchers have presentedefficient methods using data mining and artificial intelligence tools to identify potentialchurners. In order to improve the process of predicting customer churn, in this paper we propose a novelensemble learningbased approach, in which the Gray Wolf Optimization(GWO) algorithm is utilized to select the effective features and also adjust the hyper-parameters in the proposed model. We have implemented our proposed model using Python and simulated it on the IBM Telco dataset to evaluate itsperformanceand compared the obtained results with existingmethods using common evaluation criteria including accuracy, precision, recall,and AUC. The experimentalresults show the superiority of the proposed method over other evaluated solutions.

In recent years, the massive growth of generated content by users in social networks and online marketing sites, allows people to share their feelings and opinions on a variety of opinions about different products and services. Sentiment analysis is an important factor for better decision-making that is done using natural language processing (NLP), computational methods, and text analysis to extract the polarity of unstructured documents. The complexity of human languages and sentiment analysis have created a challenging research context in computer science and computational linguistics. Many researchers used supervised machine learning algorithms such as Naïve Bayes (NB), Stochastic Gradient Descent (SGD), Support Vector Machine (SVM), Logistic Regression (LR) Random Forest (RF), and deep learning algorithms such as Convolution Neural Network (CNN) and Long Short-Term Memory (LSTM). Some researchers have used Dictionary-based methods. Despite the existence of effective techniques in text mining, there are still unresolved challenges. Note that user comments are unstructured texts; Therefore, in order to structure the textual inputs, parsing is usually done along with adding some features, linguistic interpretations and removing additional items, and inserting the next terms in the database, then extracting the patterns in the structured data and finally the outputs will evaluate and interpret. The imbalance of data with the difference in the number of samples in each class of a dataset is an important challenge in the learning phase. This phenomenon breaks the performance of the classifications because the machine does not learn the features of the unpopulated classes well. In this paper, words are weighted based on the prescribed dictionary to influence the most important words on the result of the opinion mining by giving higher weight. On the other hand, the combination of the adjacent words using n-gram methods will improve the outcome. The dictionaries are highly related to the domain of the application. Some words in an application are important but in mobile comments are not impressive. Another challenge is the unbalanced train data, in which the number of positive sentences is not equal to the number of negative sentences. In this paper, two ideas are applied to build an efficient opinion mining algorithm. First, we build a precise dictionary for mobile Persian comments, and the second idea is to balance the positive and negative comments in train data. In summary, the main achievements of the current research can be mentioned: creating a weighted comprehensive dictionary in the field of mobile phone opinions to increase the accuracy of opinion analysis, balancing positive and negative opinions to improve the accuracy of opinion analysis, and eliminating the negative effect of overfitting and providing a precise approach to Determining the polarity of users' opinions about mobile phones using machine learning and recurrent deep learning algorithms. This new method is presented on mobile phone products from the Digikala site and Senti-Pers data. The result is performed with Naive Bayesian, Support Vector Machine, Stochastic Gradient Descent, Logistic Regression, Random Forest, and deep learning methods such as Convolutional Neural Network and Long Short-Term Memory based on parameters such as Accuracy, Precision, Retrieval, and F-Measure. The proposed method increases accuracy on Digikala, with NB between 10% and 34% and SVM between 5% and 24%, SGD between 7% and 38%, LR between 5% to 38%, and RF between 4% Up to 22% and CNN by 4%. The results show an accuracy increment on Senti-Pers, with NB between 12% and 46% and SVM between 5% and 46%, SGD between 5% and 35%, LR between 6% to 46%, and RF between 4% Up to 46%.

Recommender systems that predict users' ratings for a set of fonts are known as a subset of information filtering systems. Movie recommendation systems are one of the most well-known and widely used systems that have been discussed in this research. Movie recommendation systems are efficient tools that help users find their favorite movies by checking users' previous interests. These systems are based on users' ratings of past movies and use them to predict their interests in the future. However, the inappropriate scoring that users provide leads to a problem called data sparsity. This problem reduces the efficiency of movie recommendation systems.On the other hand, other available data, such as the genre of movies and demographic information of users, play a vital role in helping recommender methods to generate better recommendations. This paper proposes a movie recommendation method using movie genres and users' demographic information. Also, we propose an efficient model to evaluate the user's scoring profile and determine the minimum score required to produce an accurate prediction.Then, suitable virtual scores are combined with profiles that have inappropriate scores. These virtual scores are calculated using similarity values ​​between users obtained from movie genres and users' demographic information. In addition, a useful criterion for determining the reliability of an item is introduced, which ensures the reliability of virtual scoring. Finally, unknown scores are predicted for the target user based on the developed scoring profiles. Experiments conducted on two famous movie recommendation datasets show that the proposed approach is more efficient than other compared recommenders.

Fast and accurate answer to the questions asked in natural language is one of the important goals in the development of question and answer systems in which the computer understands a context and question and provides the exact answer to the user. Although there has been a lot of progress in this area, it is still among the issues that need to be improved, especially for languages ​​other than English, such as Persian. FarsiQuAD (FarsiQuAD) was created by humans from Persian Wikipedia articles and published in two versions. Version 1 contains 10,000+ questions and answers and version 2 contains a collection of over 145,000+ rows. This database has the ability to integrate with the English version of SQuAD and other databases of other languages ​​that have used this standard.The results of this research show that the created Persian language question and answer database can provide the user with the answer to the questions asked in the natural Persian language with an exact matching criterion of 78%  and an F1 criterion of 87%, and it still needs to be improved.

Natural language processing is growing significantly and has gained much attention with the advent of the World Wide Web and search engines, and researchers have witnessed an explosion of information in different languages. Sentiment analysis is one of the most active fields of study in natural language processing that focuses on text classification and is used to identify, extract and analyze subjective information from text sources. Aspect-based sentiment analysis is a text analysis technique that classifies comments by aspect and identifies the sentiment associated with each aspect. This analysis can be used to automatically analyze the feedback of customers' comments to different parts of goods or services and help employers to focus on points that need quality improvement. In this paper, we will introduce a new architecture based on deep learning for aspect-based sentiment analysis. This architecture will use an attention-encoder network-based model with multiple multi-head attention and a pointwise convolutional transform (which is a parallelizable and interactive alternative to LSTM and is applied to compute hidden states of input embeddings). Testing this architecture on three different datasets, including restaurants and laptops, SemEval 2014 Task 4 and  ACL 14 Twitter dataset, in all three datasets, the polarity of emotions is positive, neutral and negative, which is compared with modern methods of sentiment analysis. Based on the aspect, it will show the high accuracy of this method. For example, the aspect-based sentiment analysis on the restaurant dataset has shown 0.8187% accuracy, which has increased the accuracy by 1.97% compared to modern methods.

Computer games have played an important role in the development of artificial intelligence in recent years. Throughout the history of artificial intelligence, computer games have been a suitable test environment for evaluating new approaches and algorithms to artificial intelligence. Different methods, including rule-based methods, tree search methods, and machine learning methods (supervised learning and reinforcement learning) have been developed to create intelligent agents in different games. Games have been used as a suitable environment for trial and error, testing different artificial intelligence ideas and algorithms. Among these researches, we can mention the research of Deep Blue in the game chess and AlphaGo in the game Go. AlphaGo is the first computer program to defeat an expert human Go player. Also, Deep Blue is a chess-playing expert system is the first computer program to win a match, against a world champion. In this paper, we focus on the match-3 game. The match-3 game is a popular game in cell phones, which consists of a very large random state space that makes learning difficult. It also has random reward function which makes learning unstable.  Many researches have been done in the past on different games, including match-3. The aim of these researches has generally been to play optimally or to predict the difficulty of stages designed for human players. Predicting the difficulty of stages helps game developers to improve the quality of their games and provide a better experience for users. Based on the approach used, past works can be divided into three main categories including search-based methods, machine learning methods and heuristic methods. In this paper, an intelligent agent based on deep reinforcement learning is presented, whose goal is to maximize the score in the match-3 game. Reinforcement learning is one of the approaches that has received a lot of attention recently. Reinforcement learning is one of the branches of machine learning in which the agent learns the optimal policy for choosing actions in different spaces through its experiences of interacting with the environment. In deep reinforcement learning, reinforcement learning algorithms are used along with deep neural networks. In the proposed method, different mapping mechanisms for action space and state space are used. Also, a novel structure of neural network for the match-3 game environment has been proposed to achieve the ability to learn large state space. The contributions of this article can be summarized as follow. An approach for mapping the action space to a two-dimensional matrix is presented in which it is possible to easily separate valid and invalid actions. An approach has been designed to map the state space to the input of the deep neural network, which reduces the input space by reducing the depth of the convolutional filter and thus improves the learning process. The reward function has made the learning process stable by separating random rewards from deterministic rewards. The comparison of the proposed method with other existing methods, including PPO, DQN, A3C, greedy method and human agents shows the superior performance of the proposed method in the match-3 game

Reinforcement learning (RL) refers to the training of machine learning models to make a sequence of decisions on which an agent learns by interacting with its environment, observing the results of interactions and receiving a positive or negative reward, accordingly. RL has many applications for multi-agent systems, especially in dynamic and unknown environments. However, most multi-agent reinforcement learning (MARL) algorithms suffer from some problems specifically the exponential computational complexity to calculate the joint state-action space, which leads to the lack of scalability of algorithms in realistic multi-agent problems. Applications of MARL can be categorized from robot soccer, networks, cloud computing, job scheduling, and to optimal reactive power dispatch. In the area of reinforcement learning algorithms, there are serious challenges such as the lack of application of equilibrium-based algorithms in practice and high computational complexity to find equilibrium.  On the other hand, since agents have no concept of equilibrium policies, they tend to act aggressively toward their goals, which it results the high probability of collisions. Consequently, in this paper, a novel algorithm called Regularized Knowledge Transfer for Multi-Agent Reinforcement Learning (RKT-MARL) is presented that relies on Markov decision process (MDP) model. RKT-MARL unlike the traditional reinforcement learning methods exploits the sparse interactions and knowledge transfer to achieve an equilibrium across agents. Moreover, RKT-MARL benefits from negotiation to find the equilibrium set. RKT-MARL uses the minimum variance method to select the best action in the equilibrium set, and transfers the knowledge of state-action values across various agents. Also, RKT-MARL initializes the Q-values in coordinate states as coefficients of current environmental information and previous knowledge. In order to evaluate the performance of our proposed method, groups of experiments are conducted on five grid world games and the results show the fast convergence and high scalability of RKT-MARL. Therefore, the fast convergence of our proposed method indicates that the agents quickly solve the problem of reinforcement learning and approach to their goal.