Mohaghegh, S. Noferesti*, and M. Rajaei Abstract: In the era of big data, automatic data analysis techniques such as data mining have been widely used for decision-making and have become very effective. Among data mining techniques, classification is a common method for decision making and prediction. Classification algorithms usually work well on balanced datasets. However, one of the challenges of the classification algorithms is how to correctly predicting the label of new samples based on learning on imbalanced datasets. In this type of dataset, the heterogeneous distribution of the data in different classes causes examples of the minority class to be ignored in the learning process, while this class is more important in some prediction problems. To deal with this issue, in this paper, an efficient method for balancing the imbalanced dataset is presented, which improves the accuracy of the machine learning algorithms to correct prediction of the class label of new samples. According to the evaluations, the proposed method has a better performance compared to other methods based on two common criteria in evaluating the classification of imbalanced datasets, namely "Balanced Accuracy" and "Specificity".

The purpose of stance detection is to identify the author's stance toward a particular topic or claim. Stance detection has become a key component in applications such as fake news detection, claim validation, argument searching, and author profiling. Although significant progress has been made in stance detection in languages such as English, little attention has been paid in some other languages, including Persian.  One of the main problems of research in Persian stance detection is the shortage of appropriate datasets. In this article, to address this problem, we consider data augmentation, the artificial creation of training data, which is used to conquer the shortage of datasets. In this research, we studied several methods of data augmentation such as EDA, back-translation, and merging source dataset with similar one in English language. The experimental results indicate that combining the primary data set with the translation of another dataset with similar content in another language (for example English) result in a significant improvement in the performance of the model.

The purpose of speech emotion recognition systems is to create an emotional connection between humans and machine, since recognizing human emotions and goals helps improve interactions between humans and machines. Recognizing emotions through speech has been a challenge for researchers over the past decade. But with advances in artificial intelligence, these challenges have faded. In this study, we took steps to improve the efficiency of these systems by using deep learning methods. In the first step, three-dimensional Convolutional neural networks are used to learn the spectral-temporal Features of speech. In the second step, to strengthen the proposed model, We use the New pyramidal Concatenated three-dimensional Convolutional neural networks, Which is a multi-scale architecture of three-dimensional Convolutional neural networks on input dimensions. Finally, to obtain the ability of learning the spectral-temporal features extracted from the New Pyramidal Concatenated 3D CNN Approach, we used the temporal capsule network, so could be called consider the spatial and temporal relationship of the data. Finally, we named the proposed structure, which is a powerful structure for spectral-temporal feaures, the MSID 3DCNN + Temporal Capsule.The final model has been applied on a combination of two speech and song databases from the RAVDESS database. comparing the results of the proposed model with the conventional models, shows the better performance of our approach. The proposed SER model has achieved an accuracy of 81.77% for six emotional classes by gender.

One of the common methods for monitoring fetal growth is measuring its head circumference in ultrasound images taken from the mother's womb. In recent years utilizing deep learning methods have been expanded in this application thanks to its potential in promoting the accuracy of estimating head circumference. However, the performance of deep neural networks is highly dependent on the volume of training data. On the other hand, the region of the fetal head is segmented with considerable errors, due to the presence of various types of noise. In this article, a new method is presented to improve fetal head circumference estimation in ultrasound images in which by using unsupervised data augmentation an attempt is made to increase the amount of training data of the deep network. Parallelly by utilizing an elliptical contour estimation method, an optimal contour is created to decrease the segmentation errors . Comparing the performance of the proposed scheme with the basic method as well as state-of-art schemes shows the improvement of fetal head circumference estimation with the help of the proposed algorithm in such way that not only the quality of fetal head circumference measurement with the Dice parameter has been improved by 0.6% and 3.24% respectively compared to the closest alternative and the basic method, but also the variance of the obtained results in both types of these comparisons have improved dramatically. These achievements demonstrate the performance of the proposed method is also more focused and reliable in addition to being more accurate.

Adequate training data is essential in all supervised learning methods, including deep learning and machine vision. One of the approaches used to increase the number of training examples in deep learning is the "data augmentation" method. This method involves rotation transformation, transitions, and cropping on training images, which leads to an increase in the number of samples, which are different from training data. In this paper, the "style transfer" algorithm is used to increase the number of training samples. The goal in style transfer is to apply the appearance or visual style of one image to another image. In this paper, this method is used to produce new training examples and as an application, the proposed method is applied to the problem of fire detection. Assuming that the training images recorded during the night are less than the samples taken during the day, by applying a style transfer method, the images of the day are converted into night images and added to the data set as training data. The test results show the efficiency of the proposed data augmentation method. On average, the correct detection rate has increased by 7%.

Accurate and timely diagnosis of liver lesions in medical imaging remains a fundamental challenge in healthcare, requiring advanced techniques and sufficient datasets to enhance diagnostic performance. This study proposes a hybrid approach to improve liver lesion classification in CT images by integrating synthetic data generation through Generative Adversarial Networks (GANs) with classical data augmentation methods such as rotation, flipping, and scaling. Initially, conventional augmentation techniques were employed to expand the training dataset, followed by the generation of high-quality synthetic images using GANs. Experimental results demonstrated that the combined use of these datasets improved the model's sensitivity from 77. 5% to 85% and diagnostic accuracy from 87. 3% to 93. 1%. By highlighting the critical role of synthetic data, this research takes a significant step toward enhancing the performance of deep learning models in the automated detection of liver lesions. Furthermore, the findings underscore the potential of AI-driven techniques to facilitate the development of innovative diagnostic tools, reducing costs and human errors in medical processes.