deep learning

For investors, predicting stock market changes has always been attractive and challenging because it helps them accurately identify profits and reduce potential risks. Deep learning-based models, as a subset of machine learning, receive attention in the field of price prediction through the improvement of traditional neural network models. In this paper, we propose a model for predicting stock prices of Tehran Stock Exchange companies using a long-short-term memory (LSTM) deep neural network. The model consists of two LSTM layers, one Dense layer, and two DropOut layers. In this study, using our studies and evaluations, the adjusted stock price with 12 technical index variables was taken as an input for the model. In assessing the model's predictive outcomes, we considered RMSE, MAE, and MAPE as criteria. According to the results, integrating technical indicators increases the model's accuracy in predicting the stock price, with the LSTM model outperforming the RNN model in this task.

The capital market plays a crucial role within a country's financial structure and is instrumental in funding significant, long-term projects. Investments in the railway transport industry are vital for boosting other economic areas and have a profound impact on macroeconomic dynamics. Nonetheless, the potential for delayed or uncertain returns may deter investors. Accurate predictions of rail company stock prices on exchanges are therefore vital for making informed investment choices and securing sustained investment. This study employs deep learning techniques to forecast the closing prices of MAPNA and Toucaril shares on the Tehran Stock Exchange. It utilizes deep neural networks, specifically One-dimensional Convolutional Neural Networks (1D-CNN), Long Short-Term Memory (LSTM) networks, and a combined CNN-LSTM model, for stock price prediction. The effectiveness of these models is measured using various metrics, including MAE, MSE, RMSE, MAPE, and R2. Findings indicate that deep learning methods can predict stock prices effectively, with the CNN-LSTM model outperforming others in this research. According to the results, The CNN-LSTM model reached the highest R2 of 0.992. Also, based on criteria such as MAE, MSE, RMSE, and MAPE the best results belong to LSTM (Kaggle-modified) with 521.715, 651119.194, 806.920, and 0.028, respectively.

Lung infection represents one of the most perilous indicators of Covid-19. The most efficient diagnostic approach entails the analysis of CT scan images. Utilizing deep learning algorithms and machine vision, computer scientists have devised a method for automated detection of this disease. This study proposes a two-stage approach to identifying lung infection. In the initial stage, image features are extracted through a transfer learning framework employing ResNet50, with the last two layers being fixed. Subsequently, a CNN neural network is constructed for image detection and categorization in the second stage. By employing superior image feature selection and minimizing non-informative features, this proposed method achieves impressive accuracy metrics: 98.99% accuracy, 98.91% sensitivity, and 99.10% specificity. Furthermore, a comparative analysis is conducted between this method and six other architectures (Inception, InceptionResNetV2, ResNet101, ResNet152, VGG16, VGG19), with and without transfer learning. The findings demonstrate that the proposed method attains 98% accuracy on test data, without succumbing to overfitting.

Over the past few years, there has been a significant increase in patent applications, which has resulted in a heavier workload for examination offices in examining and prosecuting these inventions. To adequately perform this legal process, examiners must thoroughly analyze patents by manually identifying the semantic information such as problem description and solutions. The process of manually annotating is both tedious and time-consuming. To solve this issue, we have introduced a deep ensemble model for semantic paragraph-level pattern classification based on the semantic content of patents. Specifically, our proposed model classifies the paragraphs into semantic categories to facilitate the annotation process. The proposed model employs stack generalization as an ensemble method for combining various deep models such as Long Short-Term Memories (LSTM), bidirectional LSTM (BiLSTM), Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), and the pre-trained BERT model. We compared the proposed model with several baselines and state-of-the-art deep models on the PaSA dataset containing 150000 USPTO patents classified into three classes of 'technical advantages', 'technical problems', and 'other boilerplate text'. The results of extensive experiments show that the proposed model outperforms both traditional and state-of-the-art deep models significantly.

According to numerous studies, the Qur’anic surahs exhibit internal structure and organization, with each surah serving a distinct purpose. Although each surah focuses on a specific theme and the Qur’an identifies 114 broad themes, the arrangement of the surahs and the remarkable similarity between adjacent surahs (neighbors) underscores the chain-link and deliberate positioning of the surahs within the Qur’an. To investigate this phenomenon, a multifaceted and compound model was developed, comprising two main parts: embedding and autoencoding. The first part was carried out by preparing the words and roots of the Qur’anic text using the BERT model for meaning-topic representation. In the second part, the data was clustered in a soft labeling mode by the autoencoder. Analysis of the distribution of surahs within clusters revealed that neighboring surahs exhibited an average similarity of 80, while surahs with greater distance showed an average similarity of 20. The findings support the placement of similar surahs in close proximity,  substantiating the organized sequence of Qur’anic surahs. To conclude, the results provide compelling evidence for the structured arrangement of Qur’anic surahs.

Cyberbullying involves using hurtful or offensive language that goes against basic rules of respect and politeness. It harms the online environment and can negatively affect people by causing harassment, discrimination, or emotional pain. To combat this, it is crucial to develop automated methods for detecting and preventing the dissemination of such content. Deep learning, a branch of artificial intelligence, leverages neural networks to learn from data and perform complex tasks, effectively capturing semantic and grammatical nuances to differentiate between abusive and non-abusive language. This survey paper reviews current techniques and advancements in deep learning-based approaches for detecting cyberbullying content on online platforms, aiming to provide a comprehensive understanding of existing methodologies and identify potential avenues for future research to mitigate the spread and impact of such behaviors on the internet.

The term "clickbait" refers to content specifically designed to capture readers' attention, often through misleading headlines, leading to frustration among social media users. In this study, titled "Mushakkal," which translates to "variety" in Arabic, we utilized a Convolutional Neural Network (CNN)—a deep learning approach—to detect clickbait within an Arabic dataset. We compared three optimizers: RMSprop, Adam, and Adadelta, evaluating various parameter settings to determine the most effective combination for detecting clickbait in Arabic content. Our findings revealed that the CNN model performed best when both pre-processing and Word2Vec techniques were applied. The Adam optimizer outperformed the others, achieving a Macro-F1 score of 77%. The RMSprop optimizer closely followed, attaining a Macro-F1 score of 76%. In contrast, Adadelta proved to be the least effective for classifying Arabic text.

Facial emotion recognition has recently attracted considerable interest due to its wide range of applications. It plays a crucial role in supporting individuals with autism spectrum disorders and improving interactions between humans and computers. The ability to execute these applications in real-time is essential. The architecture of the model and the computational resources available are the key determinants of inference time. Consequently, the development of a real-time solution requires a concentrated effort on these elements. In this paper, we present a scalable approach that utilizes EfficientNetV2, chosen for its operational efficiency. Our methodology involves resolution scaling based on a polynomial equation, which ensures real-time performance across various computational resources and model configurations. This scalable technique employs a polynomial equation to identify the optimal resolution for designated inference times, specifically adapted to our hardware and model specifications. By implementing the polynomial equation for resolution scaling, we created two variants of EfficientNetV2. Our findings from the KDEF dataset indicate that the proposed EfficientNetV2 can accurately classify images in real time on our hardware.

With the expansion of smart homes, Human Activity Recognition (HAR) has become a key challenge in artificial intelligence, enhancing not only the comfort and safety of residents but also contributing to the development of applications such as healthcare and smart surveillance. The Transformer architecture, with its ability to model long-term dependencies and process data in parallel, has made significant advancements in recognizing human activities. In addition, its multi-head attention mechanism enables the analysis of complex input data by allowing the model to focus on different parts of the input simultaneously, capturing diverse relationships and dependencies within the data. This paper examines the application of Transformers in HAR and analyzes recent studies (since 2019). In addition to investigating innovative architectures, feature extraction methods, and accuracy improvements, it also discusses the challenges and future prospects of these models in recognizing human activities. Rapid advancements in deep learning and access to extensive datasets have made Transformers a key tool for improving the accuracy and efficiency of HAR systems in smart environments.