In recent years, deep learning neural networks have been recognized as powerful tools for solving complex problems. Deep learning is a subfield of artificial intelligence in which complex problems with numerous parameters and inputs are modeled based on a set of algorithms. In this research, a new framework of deep learning is presented. Using wavelet transform, stacked auto-encoders, and the Long Short-Term Memory or LSTM, we predict the market direction in the future contracts of gold coins of Iran's Commodity Exchange market. The input data is first denoised using the wavelet transformer in the proposed method. Then, using the stacked auto-encoder, the indicators influencing the market direction are identified. Ultimately, these indicators are given as input to the LSTM architecture to predict the market direction. Proposing several new technical indicators to increase the accuracy of the proposed model, adjusting the parameters of the utilized algorithms, including LSTM, for this problem, and suggesting a trading strategy to achieve appropriate profitability are among the contributions of the present study. Investigations reveal that the proposed method outperforms other approaches and achieves higher accuracy and efficiency.

The development of new technologies has confronted the entire domain of science and industry with issues of big data's scalability as well as its integration with the purpose of forecasting analytics in its life cycle. In predictive analytics, the forecast of near-future and recent past - or in other words, the now-casting - is the continuous study of real-time events and constantly updated where it considers eventuality. So, it is necessary to consider the highly data-driven technologies and to use new methods of analysis, like machine learning and visualization tools, with the ability of interaction and connection to different data resources with varieties of data regarding the type of big data aimed at reducing the risks of policy-making institution’s investment in the field of IT. The main scientific contribution of this article is presenting a new approach of policy-making for the now-casting of economic indicators in order to improve the performance of forecasting through the combination of deep nets and deep learning methods in the data and features representation. In this regard, a net under the title of P-V-L Deep: Predictive Variational Auto Encoders - Long Short-term Memory Deep Neural Network was designed in which the architecture of variational auto-encoder was used for unsupervised learning, data representation, and data reconstruction; moreover, long short-term memory was adopted in order to evaluate now-casting performance of deep nets in time-series of macro-econometric variations. Represented and reconstructed data in the generative network of variational auto-encoder to determine the performance of long-short-term memory in the forecasting of the economic indicators were compared to principal data of the net. The findings of the research argue that reconstructed data which are derived from variational auto-encoder embody shorter training time and outperform of prediction in long short-term memory compared to principal data.

The huge amount of information has forced researchers to find a solution to face this fundamental problem called data overload. Recommender systems try to suggest the required information to the user by examining the user's preferences directly or based on the behavior of other similar users in a way that best matches the user's needs. Meanwhile, the use of textual information hidden in the user's biography or comments can be very useful. Declarative systems try to find similarities by examining each word in users' comments with the comments of other users, this is if different meanings for a word are ignored. In this way, the use of auto-encoder networks in order to check the semantic relationship of words in a sentence with respect to the opinions of other users can overcome this challenge. In this article, a personalization approach is presented based on the recommendation system in social networks using the combination of collaborative filter and deep auto-encoder networks. In proposed recommendation system, the information in the user profile and user comments to each website is used as the input of the presented combined deep auto-encoder network and the collaborative filter method in order to find similar users accurately and predict the website's rating by users. Finally, after finding similar users, it provides recommendations to visit and personalize the web page of serious users based on the favorite websites of similar users. Due to the convolutional layers of proposed deep auto-encoder network, the training process in the middle layer has performed on semantic relationship of words in a sentence to find similar comments and users. This method implemented on two standard datasets titled TripAdvisor and Yelp. The proposed method has improved in terms of statistical accuracy of about 39%, the ratio of successful recommendations to useful recommendations of about 6%, and the accuracy of recognizing similar users is about 18% from other classification methods.

In network analysis, a community is typically considered of as a group of nodes with a great density of edges among themselves and a low density of edges relative to other network parts. Detecting a community structure is important in any network analysis task, especially for revealing patterns between specified nodes. There is a variety of approaches presented in the literature for overlapping and disjoint detection of community in networks. In recent years, many researchers have concentrated on feature learning and network embedding methods for node clustering. These methods map the network into a lower-dimensional representation space. We propose a model in this research for learning graph representation using deep neural networks. In this method, a nonlinear embedding of the original graph is fed to stacked auto-encoders for learning the model. Then an overlapping clustering algorithm is performed to obtain overlapping communities. The effectiveness of the proposed model is investigated by conducting experiments on standard benchmarks and real-world datasets of varying sizes. Empirical results exhibit that the presented method outperforms some popular community detection methods.

Recent researches have determined that regularized auto-encoders can provide a good representation of data which improves the performance of data classification. These type of auto-encoders which are usually over-complete, provide a representation of data that has some degree of sparsity and is robust against variation of data to extract meaningful information and reveal the underlying structure of data by making a change in classic auto- encoders’ structure and/or adding regularizing terms to the objective function. The present study aimed to propose a novel approach to generate sparse, robust, and discriminative features through supervised regularized auto-encoders, in which unlike most existing auto-encoders, the data labels are used during feature extraction to improve discrimination of the representation and also, the sparsity ratio of the representation is completely adaptive and dynamically determined based on data distribution and complexity. Results reveal that this method has better performance in comparison to other regularized auto-encoders regarding data classification.

Diagnosing the type of cancer, which is called the subtype, is very important in determining the treatment process. This paper focuses on the diagnose of the four subtypes of the brain cancer. Disease subtype diagnosis can be modeled as a classification problem. Due to the significant progress made in bioinformatics in extracting genetic information from the human body, recently this information is widely used in the representing of patients in machine learning. In this paper, three types of genetic information including mRNA, miRNA and DNA methylation are used.It should be noted that combining different information sources in the form of multimodal data instead of using a single information source increases the accuracy of information classification. To extract more desirable features from the original genetic data, auto-encoder has been used so that the features extracted from auto-encoder are concatenated to the original genetic data.Random forest has performed well as a classifier in classifying patients based on genetic information. By extending deep methods in neural networks and their good performance, a version of deep random forest with layered structure has been proposed. The deep random forest has the advantage that has a limited number of parameters and lower computational complexity in addition to the optimal performance in information classification. In this paper, deep random forest is used to determine the subtype of a special type of brain cancer. The experiment results show the desired performance of the proposed method.

Community detection is an important topic for social network analysis and is also essential to understanding complex networks structure. In community detection, the goal is to determine the groups in which the group nodes are densely connected to each other. In this research, deep learning techniques have been used to control graph data with high dimensions, while presenting a comprehensive and integrated architecture of community recognition methods with deep learning. Community detection classic approaches are suitable for networks with low dimensions. Therefore, the reduction of complex network dimensions is counted as a significant topic in community detection. In this paper, in order to reveal the direct and indirect connections among nodes, first a new similarity matrix of network topology is built. Then, a stacked auto-encoder is designed to decrease dimensions based on unsupervised learning. In order to detect communities, various clustering algorithms are then tested and utilized. Evaluation of the proposed research model is performed by surveying various experiments on standard criteria and six real data sets of Karate, Dolphins, Football, Polbooks, Cora and Citeseer. The proposed method evaluation outcomes show a higher accuracy in the identification of communities in the football data set compared to the twelve proposed algorithms used in past researches, and show a significant improvement in other data sets compared to the thirteen algorithms.

Identification of promising areas associated with mineralization and integration of exploratory multi-resource data-sets are essential in mineral potential modeling. In this research, big data analysis method and an unsupervised deep auto-encoder network algorithm were used to identify the exploratory targets areas associated with porphyry copper-gold mineralization in the Dehsalm strict of Iran. The results show that the identified exploratory target areas have strong spatial relationships with known mineral indices in the study area. The Prediction-Area (P_A) plot analysis shows that the generated model performs well. The result of this study demonstrates that big data analytics supported by deep learning methods is a potential technique to be considered for use in mineral prospectivity mapping.

Web Application Firewall (WAF) is known as one of the Intrusion Detection System (IDS) solutions for protecting web servers from HTTP attacks. WAF is a tool to identify and prevent many types of attacks, such as XSS and SQL-injection. In this paper, deep machine learning algorithms are used for enriching the WAF based on the anomaly detection method. Firstly, we construct attributes from HTTP data, to do so we consider two models namely n-gram and one-hot. Then, according to Auto-Encoder LSTM (AE-LSTM) as an unsupervised deep leaning method, we should extract informative features and then reduce them. Finally, we use ensemble isolation forest to train only normal data for the classifier. We apply the proposed model on CSIC 2010 and ECML/ PKDD 2007 datasets. The results show AE-LSTM has higher performance in terms of accuracy and generalization compared with naïve methods on CSIC dataset; the proposed method also have acceptable detection rate on ECML/PKDD dataset using n -gram model.

Interoceptions are a combination of sensation, integration, and interpretation of internal bodily signals. Interoceptions are bidirectionally related to the human being mental and physiological health, and well-being. Sleep and different interoceptive modalities are proven to share common relations.Heartbeat Evoked Potential (HEP) is known as a robust readout to interoceptive processes. In this study, we focused on the relation between HEP modulations and sleep-related disorders.

We investigated four different sleep-related disorders, including insomnia, rapid eye movement behavior disorder, periodic limb movements and nocturnal frontal lobe epilepsy, and provided HEP signals of multiple Electroencephalogram (EEG) channels over the right hemisphere to compare these disorders with the control group. Here, we investigated and compared the results of 35 subjects, including seven subjects for the control group and seven subjects for each of above-mentioned sleep disorders.

By comparing HEP responses of the control group with sleep-related patients’ groups, statistically significant HEP differences were detected over right hemisphere EEG channels, including FP2, F4, C4, P4, and O2 channels. These significant differences were also observed over the grand average HEP amplitude activity of channels over the right hemisphere in the sleep-related disorders as well.

Our results between the control group and groups of patients suffering from sleep-related disorders demonstrated that during different stages of sleep, HEPs show significant differences over multiple right hemisphere EEG channels. Interestingly, by comparing different sleep disorders with each other, we observed that each of these HEP differences’ patterns over specific channels and during certain sleep stages bear considerable resemblances to each other.