Search

Abstract Credit scoring models are critical for financial institutions to assess borrower risk and maintain profitability. Although machine learning models have improved credit scoring accuracy, imbalanced class distributions remain a major challenge. The widely used Synthetic Minority Oversampling TEchnique (SMOTE) struggles with high-dimensional, non-linear data and may introduce noise through class overlap. Generative Adversarial Networks (GANs) have emerged as an alternative, offering the ability to model complex data distributions. Conditional Wasserstein GANs (cWGANs) have shown promise in handling both numerical and categorical features in credit scoring datasets. However, research on extracting latent features from non-linear data and improving model explainability remains limited. To address these challenges, this paper introduces the Non-parametric Oversampling Technique for Explainable credit scoring (NOTE). The NOTE offers a unified approach that integrates a Non-parametric Stacked Autoencoder (NSA) for capturing non-linear latent features, cWGAN for oversampling the minority class, and a classification process designed to enhance explainability. The experimental results demonstrate that NOTE surpasses state-of-the-art oversampling techniques by improving classification accuracy and model stability, particularly in non-linear and imbalanced credit scoring datasets, while also enhancing the explainability of the results.

Fault detection in early operational stages of rolling bearing is crucial for reliable and safe functioning of rotating machinery. Implementation of intelligent fault detection techniques involving deep learning methods enable automatic feature extraction and selection from raw vibration data provides accurate results. The shortage of enough historical data limits the application of deep learning. Therefore, to solve this problem, in this paper data augmentation method is implemented to generate new data that having greater similitude with the real data for better training of deep learning model for fault detection. For this purpose, WGAN (Wasserstein generative adversarial network) is implemented as imbalanced data augmentation method. Also SAE (stacked autoencoder) is implemented to obtain the latent representation of raw vibration data which is used as noise vector to train WGAN. This has greatly improved the quality of data generation from WGAN. The quality assessment of generated samples is quantified by implementing metrics such as KLD (Kullback-Leibler divergence) and NCC (normalized cross-correlation). The comparison with conventional data generation methods such as VAE, and GAN proves the superior quality of data generation by SAE-WGAN. Test rig experiments are used to gather the vibration data, and deep convolutional neural networks are used to classify the faults (DCNN). The ROC (receiver operating characteristic) curve and performance evaluation metrics like precision, recall, and F1-score amply demonstrated the excellent discriminative power of the suggested methodology for fault detection. Hence the proposed work successfully implemented as condition monitoring tool for early fault detection in rotating machinery. [ABSTRACT FROM AUTHOR]

Credit scoring models are critical for financial institutions to assess borrower risk and maintain profitability. Although machine learning models have improved credit scoring accuracy, imbalanced class distributions remain a major challenge. The widely used Synthetic Minority Oversampling TEchnique (SMOTE) struggles with high-dimensional, non-linear data and may introduce noise through class overlap. Generative Adversarial Networks (GANs) have emerged as an alternative, offering the ability to model complex data distributions. Conditional Wasserstein GANs (cWGANs) have shown promise in handling both numerical and categorical features in credit scoring datasets. However, research on extracting latent features from non-linear data and improving model explainability remains limited. To address these challenges, this paper introduces the Non-parametric Oversampling Technique for Explainable credit scoring (NOTE). The NOTE offers a unified approach that integrates a Non-parametric Stacked Autoencoder (NSA) for capturing non-linear latent features, cWGAN for oversampling the minority class, and a classification process designed to enhance explainability. The experimental results demonstrate that NOTE surpasses state-of-the-art oversampling techniques by improving classification accuracy and model stability, particularly in non-linear and imbalanced credit scoring datasets, while also enhancing the explainability of the results. [ABSTRACT FROM AUTHOR]

A concurrent heuristic search iterative process (CHSIP) is used for estimating groundwater pollution sources and aquifer parameters in this work. Frequent calls to carry out a numerical simulation of groundwater pollution have generated a huge calculated load during the CHSIP. Therefore, a valid means to mitigate this is building a substitute to emulate the numerical simulation at a low calculated load. However, there is a complicated nonlinear correlativity between the import and export of the numerical simulation on account of the large quantity of variables. This leads to a poor approach accuracy of the substitute compared to the simulation when using shallow learning methods. Therefore, we first built a stacked autoencoder substitute, using the deep learning method, to boost the approach accuracy of the substitute compared to the numerical simulation. In total, 400 training samples and 100 testing samples for the substitute were collected by employing the Latin hypercube sampling method and running the numerical simulator. The CHSIP was then employed for estimating the groundwater pollution sources and aquifer parameters, and the estimated outcome was obtained when the CHSIP was terminated. The data analysis, including interval estimation and point estimation, was implemented on the MATLAB platform. A relevant hypothetical case is set to verify our approaches, which shows that the CHSIP is helpful for estimating the groundwater pollution source and aquifer parameters and that the stacked autoencoder method can effectively boost the approach precision of the substitute for the simulator. [ABSTRACT FROM AUTHOR]

Diabetes is a prevalent global disease that significantly diminishes the quality of life and can even lead to fatalities due to its complications. Early detection and treatment of diabetes are crucial for mitigating and averting associated risks. This study aims to facilitate the prompt and straightforward diagnosis of individuals at risk of diabetes. To achieve this objective, a dataset for early stage diabetes risk prediction from the University of California Irvine (UCI) database, widely utilized in the literature, was employed. A hybrid deep learning model comprising genetic algorithm, stacked autoencoder, and Softmax classifier was developed for classification on this dataset. The performance of this model, wherein both the model architecture and all hyperparameters were specifically optimized for the given problem, was compared with commonly used methods in the literature. These methods include K-nearest neighbor, decision tree, support vector machine, and convolutional neural network, utilizing tenfold cross-validation. The results obtained with the proposed method surpassed those obtained with other methods, with higher accuracy rates than previous studies utilizing the same dataset. Furthermore, based on the study's findings, a web-based application was developed for early diabetes diagnosis. [ABSTRACT FROM AUTHOR]

In recent days, cardiovascular disease (CVD) has been considered one of the significant causes of morbidity and mortality in the world. Arrhythmia, acute myocardial infarction (AMI), and other cardiac illnesses have always been serious disorders that put patients' lives in danger. Different studies have shown that substantial features play a foremost role in enhancing the machine learning (ML) technique's performance. Moreover, the dataset was imbalanced, and the researchers could not utilize any model to deal with this issue. Thus, the objective is to discover the substantial features and solve the class imbalance problem that can boost accuracy and lead to better predicting cardiovascular patient survivors. In this proposed work, a novel SMOTE-based hybrid deep learning (SMOTE-HDL) network is proposed for predicting the patient's survivors of cardiovascular disease. At first, the data acquisition stage collects the input data from the dataset and then pre-processes using data normalization and missing value imputation. In addition, the class imbalance problem is solved by an oversampling technique named the synthetic minority oversampling technique (SMOTE). To extract the substantial features, an attention-based perceptive long short-term extraction network (APLSTEN) is employed. The process of feature selection is accomplished through an adaptive coati optimization algorithm (ACOA). Finally, the selected features are served to a Kepler-optimized deep stacked recurrent network (KDSRN) for attaining enhanced classification. The proposed SMOTE-HDL network is implemented in the Python platform using the Heat-failure-clinical-records dataset and assessed the performance in response to dissimilar evaluation criteria. The maximum prediction accuracy obtained by the proposed SMOTE-HDL is 95.52%, superior to the existing classifiers. [ABSTRACT FROM AUTHOR]

A Network Intrusion Detection System (NIDS) is frequently used for monitoring and detecting malicious activities in network traffic. A typical NIDS has four stages: a data source, data pre-processing, a decision-making technique, and a defense reaction. We have utilized both anomaly and signature based techniques to build a framework which is resilient to identifying both known and unknown attack. The incoming data packet is fed into the Stacked Autoencoder to identify whether it is a benign or malicious. If found to be malicious we extract the most relevant features from the network packet using grey wolf optimization algorithm. Then these attribute are provided to RandomForest classifier to determine if this malign attack is present in our knowledge base. If it is present we progress to identify the attack type using LightGBM classifier. If not, we term it as zero-day attack. To evaluate the usability of the proposed framework we have assessed it using two publicly available datasets namely UNSW-NB15 and CIC-IDS-2017 dataset. We have obtained an accuracy of 90.94% and 99.67% on the datasets respectively. [ABSTRACT FROM AUTHOR]

Dangerous human behavior in the driving sense may cause traffic accidents and even cause economic losses and casualties. Accurate identification of dangerous human behavior can prevent potential risks. To solve the problem of difficulty retaining the temporal characteristics of the existing data, this paper proposes a human behavior recognition model based on utilized optical flow and hybrid deep learning model-based 3D CNN-LSTM in stacked autoencoder and uses the abnormal behavior of humans in real traffic scenes to verify the proposed model. This model was tested using HMDB51 datasets and JAAD dataset and compared with the recent related works. For a quantitative test, the HMDB51 dataset was used to train and test models for human behavior. Experimental results show that the proposed model achieved good accuracy of about 86.86%, which outperforms recent works. For qualitative analysis, we depend on the initial annotations of walking movements in the JAAD dataset to streamline the annotating process to identify transitions, where we take into consideration flow direction, if it is cross-vehicle motion (to be dangerous) or if it is parallel to vehicle motion (to be of no danger). The results show that the model can effectively identify dangerous behaviors of humans and then test on the moving vehicle scene. [ABSTRACT FROM AUTHOR]

The prediction of patient survival is crucial for guiding the treatment process in healthcare. Healthcare professionals rely on analyzing patients' clinical characteristics and findings to determine treatment plans, making accurate predictions essential for efficient resource utilization and optimal patient support during recovery. In this study, a hybrid architecture combining Stacked AutoEncoders, Particle Swarm Optimization, and the Softmax Classifier was developed for predicting patient survival. The architecture was evaluated using the Haberman's Survival dataset and the Echocardiogram dataset from UCI. The results were compared with several Machine Learning methods, including Decision Trees, K-Nearest Neighbors, Support Vector Machines, Neural Networks, Gradient Boosting, and Gradient Bagging applied to the same datasets. The findings indicate that the proposed architecture outperforms other Machine Learning methods in predicting patient survival for both datasets and surpasses the results reported in the literature for the Haberman's Survival dataset. In the light of the findings obtained, the models obtained with the proposed architecture can be used as a decision support system in determining patient care and applied methods. [ABSTRACT FROM AUTHOR]

Developing a reliable and accurate indoor localization system is a crucial step for creating a seamless and interactive user-device experience in nearly all intelligent internet of things (IIoTs) and smart applications. Indoor localization systems based on WiFi fingerprinting have been considered as a promising alternative to model-based approaches owing to their accuracy, low cost, availability, and ease of configuration. However, recent studies have revealed that in complex environments, WiFi fingerprinting techniques are faced with a lot of challenges as the coverage area increases. These challenges include fingerprint spatial uncertainty, instability in the received signal strength indicator (RSSI) and discrepancy in fingerprint distribution. Furthermore, there is frequent need for database upgrades or even recreation whenever there is a change in the architecture of the location. These challenges have questioned the robustness and efficiency of most of the existing schemes. In this paper, we present an indoor localization architecture for complex multi-building multi-floor location prediction and subsequently propose SALLoc (SAE-ALSTM Localization), a WiFi fingerprinting indoor localization scheme based on Stacked Autoencoder (SAE) and Attention-based Long Short-Time Memory (ALSTM) framework. Firstly, stratified sampling technique is used to separate validation set from the entire uneven RSSI training set which ensures that the same proportion of RSSI samples are present in both sets. Secondly, SAE is utilized to select core features and decrease the dimensions of the RSSI samples. Finally, ALSTM is trained to focus on these features to achieve robust location prediction. Extensive investigations were conducted using UJIIndoorLoc, Tampere and UTSIndoorLoc datasets, and the results obtained demonstrated the superiority of the proposed scheme in terms of prediction accuracy, robustness, and generalizations when compared to state-of-the-art methods. The mean localization error (MLE) on UJIIndoorLoc, Tampere and UTSIndoorLoc datasets are 8.28 m, 9.52 m, and 6.48 m respectively. Consequently, it can be concluded that the proposed scheme is accurate and well-suited for large-scale indoor environment location prediction.

The risk of fraudulent activity has significantly increased with the rise in digital payments. To resolve this issue there is a need for reliable real-time fraud detection technologies. This research introduced an innovative method called stacked autoencoder kernel extreme learning machine optimized by the dandelion algorithm (S-AEKELM-DA) to detect fraudulent transactions. The primary objective was to enhance the kernel extreme learning machine (KELM) performance by integrating the dandelion technique into a stacked autoencoder kernel ELM architecture. This study aimed to improve the overall effectiveness of the proposed method in fraud detection by optimizing the regularization parameter (c) and the kernel parameter (σ). To evaluate the S-AEKELM-DA approach; simulations and experiments were conducted using four credit card datasets. The results demonstrated remarkable performance, with our method achieving high accuracy, recall, precision, and F1-score in real time for detecting fraudulent transactions. These findings highlight the effectiveness and reliability of the suggested approach. By incorporating the dandelion algorithm into the S-AEKELM framework, this research advances fraud detection capabilities, thus ensuring the security of digital transactions.

Due to the growing number of data-driven approaches, especially in artificial intelligence and machine learning, extracting appropriate information from the gathered data with the best performance is a remarkable challenge. The other important aspect of this issue is storage costs. The principal component analysis (PCA) and autoencoders (AEs) are samples of the typical feature extraction methods in data science and machine learning that are widely used in various approaches. The current work integrates the advantages of AEs and PCA for presenting an online supervised feature extraction selection method. Accordingly, the desired labels for the final model are involved in the feature extraction procedure and embedded in the PCA method as well. Also, stacking the nonlinear autoencoder layers with the PCA algorithm eliminated the kernel selection of the traditional kernel PCA methods. Besides the performance improvement proved by the experimental results, the main advantage of the proposed method is that, in contrast with the traditional PCA approaches, the model has no requirement for all samples to feature extraction. As regards the previous works, the proposed method can outperform the other state-of-the-art ones in terms of accuracy and authenticity for feature extraction.

Breast cancer (BrC) stands as the most prevalent cancer affecting women globally, comprising 24.5% of all female cancer diagnoses and contributing to 15.0% of total cancer-related fatalities. The timely detection and precise categorization of breast cancer play pivotal roles in enhancing patient prognosis and treatment outcomes. The main goal is to enhance the precision of classifying mammogram images, thus offering vital support to radiology experts in diagnosing BrCs. The proposed model encompasses several pivotal stages, including pre-processing, feature extraction, segmentation, and classification. To assess the model's efficacy, we employed the INBreast dataset. During pre-processing, mammogram images were enhanced through a customized contrast-limited adaptive histogram equalization (mCLAHE) technique coupled with data augmentation. Segmentation was executed utilizing the Res-SegNet model, and feature extraction employing the VGG-19 model. The classification was conducted via a stacked autoencoder (SAE) with particle swarm optimization (PSO). Our proposed model exhibited notably high performance compared to alternative models such as CNN, Yolo-v4, and Inception-v3. The results unveiled an accuracy of 98.33%, precision of 99.39%, recall of 98.78%, specificity of 93.75%, an F1-score of 99.08%, and an MCC score of 90.04%.

Deep learning models have been widely used in hyperspectral images classification. However, the classification results are not satisfactory when the number of training samples is small. Focused on above-mentioned problem, a novel Two-stage Multi-dimensional Convolutional Stacked Autoencoder (TMC-SAE) model is proposed for hyperspectral images classification. The proposed model is composed of two sub-models SAE-1 and SAE-2. The SAE-1 is a 1D autoencoder with asymmetric structre based on full connection layers and 1D convolution layers to reduce spectral dimensionality. The SAE-2 is a hybrid autoencoder composed of 2D and 3D convolution operations to extract spectral-spatial features from the reduced dimensionality data by SAE-1. The SAE-1 is trained with raw data by unsupervised learning and the encoder of SAE-1 is employed to reduce spectral dimensionality of raw data. The data after dimension reduction is used to train the SAE-2 by unsupervised learning. The fine-tuning of SAE-2 encoder and the training of classifier are implemented simultaneously with small number of samples by supervised learning. Comparative experiments are performed on three widely used hyperspectral remote sensing data. The extensive comparative experiments demonstrate that the proposed architecture can effectively extract deep features and maintain high classification accuracy with small number of training samples. [ABSTRACT FROM AUTHOR]

Copyright of China Pulp & Paper is the property of China Pulp & Paper Magazines Publisher and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

The alleviation of traffic congestion relies on efficient traffic control and traffic guidance, which are based on real-time short-term traffic flow prediction. In this paper, the stacked autoencoder (SAE) deep learning model with powerful feature learning capability is selected to predict the traffic flow on road sections. The process of training SAE includes the pre-training phase and the fine-tuning phase, which mainly apply the BP algorithm. However, the process of training SAE is time-consuming and cannot meet the real-time performance of modern application systems. This paper proposes a parallel training strategy for the SAE prediction model based on data parallel mode. The gradient solution process in our algorithm satisfies the conditions of parallel computing, so the training process can be designed in a parallel manner. The original dataset is distributed to some computing nodes, which are work nodes. The work node is responsible for gradient calculation using the local data. The task of the sole master node is to synthesize the gradient calculation results and then broadcast the updated gradient to each work node. The simulation results show that the SAE-based prediction model achieves better results than the traditional model, and the parallel algorithm reduces the running time of training processes. [ABSTRACT FROM AUTHOR]

Breast cancer is the common cancer in women, where early detection reduces the mortality rate. The magnetic resonance imaging (MRI) images are efficient in analyzing breast cancer, but it is hard to identify the abnormalities. The manual breast cancer detection in MRI images is inefficient; therefore, a deep learning-based system is implemented in this manuscript. Initially, the visual quality improvement is done using region growing and adaptive histogram equalization (AHE), and then, the breast lesion is segmented by Otsu thresholding with morphological transform. Next, the features are extracted from the segmented lesion, and a modified fuzzy rough set technique is proposed to reduce the dimensions of the extracted features that decreases the system complexity and computational time. The active features are fed to the stacked autoencoder for classifying the benign and malignant classes. The results demonstrated that the proposed model attained 99% and 99.22% of classification accuracy on the benchmark datasets, which are higher related to the comparative classifiers: decision tree, naïve Bayes, random forest and k-nearest neighbor (KNN). The obtained results state that the proposed model superiorly screens and detects the breast lesions that assists clinicians in effective therapeutic intervention and timely treatment. [ABSTRACT FROM AUTHOR]

The overall information of a process can be obtained through global modelling, and the local information is easily ignored in the research of the industrial process monitoring of unit connection. Thus, finding the global type of faults is easy, but occurs at the expense of drowning out the local faults. The use of block modelling can highlight local information, thereby improving local fault detection capability. However, the connection information between blocks is usually ignored in block modelling, which makes finding fault classes that only affect the connection relationship between blocks difficult. A mechanistic block‐based attention mechanism stacked autoencoder (MB‐AMSAE) monitoring method is proposed in this paper. The industrial process is divided into several parts in accordance with its mechanistic relationships, and each part represents an independent block. Self‐attention is used to focus on the information of each block itself. Cross‐attention is adopted to focus on the information between blocks, and this information is fused to form new blocks. The new block is used as the feature of the original block, and the original block is reconstructed by using a stacked autoencoder. The corresponding control limit is obtained in accordance with the reconstruction ability of normal samples, and whether the working conditions are normal is judged according to the control limit. The proposed algorithm is used in numerical simulation, Tennessee‐Eastman processes, and is compared with other advanced algorithms based on its fault detection capability. Results show the effectiveness of the MB‐AMSAE algorithm in process monitoring. [ABSTRACT FROM AUTHOR]

Timely prediction of wind turbine states is valuable for reduction of potential significant losses resulting from deterioration of health condition. To enhance the accuracy of fault diagnosis and early warning, data collected from supervisory control and data acquisition (SCADA) system of wind turbines is graphically processed and used as input for a deep learning mode, which effectively reflects the correlation between the faults of different components of wind turbines and the multi-state information in SCADA data. An improved stacked autoencoder (ISAE) framework is proposed to address the issue of ineffective fault identification due to the scarcity of labeled samples for certain fault types. In the data augmentation module, synthetic samples are generated using SAE to enhance the training data. Another SAE model is trained using the augmented dataset in the data prediction module for future trend prediction. The attribute correlation information is embedded to compensate for the shortcomings of SAE in learning attribute relationships, and the optimal factor parameters are searched using the particle swarm optimization (PSO) algorithm. Finally, the state of wind turbines is predicted using a CNN-based fault diagnosis module. Experimental results demonstrate that the proposed method can effectively predict faults and identify fault types in advance, which is helpful for wind farms to take proactive measures and schedule maintenance plans to avoid significant losses. [ABSTRACT FROM AUTHOR]

The risk of fraudulent activity has significantly increased with the rise in digital payments. To resolve this issue there is a need for reliable real-time fraud detection technologies. This research introduced an innovative method called stacked autoencoder kernel extreme learning machine optimized by the dandelion algorithm (S-AEKELM-DA) to detect fraudulent transactions. The primary objective was to enhance the kernel extreme learning machine (KELM) performance by integrating the dandelion technique into a stacked autoencoder kernel ELM architecture. This study aimed to improve the overall effectiveness of the proposed method in fraud detection by optimizing the regularization parameter (c) and the kernel parameter (σ). To evaluate the S-AEKELM-DA approach; simulations and experiments were conducted using four credit card datasets. The results demonstrated remarkable performance, with our method achieving high accuracy, recall, precision, and F1-score in real time for detecting fraudulent transactions. These findings highlight the effectiveness and reliability of the suggested approach. By incorporating the dandelion algorithm into the S-AEKELM framework, this research advances fraud detection capabilities, thus ensuring the security of digital transactions. [ABSTRACT FROM AUTHOR]

The use of neural network models to monitor bearing vibration signals can easily be affected by noise, which leads to a decrease in the model test accuracy. Therefore, the existence of noise problems increases the requirements for non-linear mapping capability and robustness of deep neural network (DNN) models. In order to deal with the noise problem, the concept of qubit neurons is introduced into a deep learning stacked autoencoder (SAE) model, and a quantum stacked autoencoder (QSAE) model based on qubits and quantum gates is proposed. The properties of SAE layer-by-layer coding and the arithmetic of qubit neurons are combined in the QSAE. The quantum state signal is taken as the input signal to the encoder and the coding activation function and coding weight matrix are redefined by quantum-controlled non-gates and quantum revolving gates, so that the quantum state signal can be coded layer by layer. Experimental results show that the QSAE can train and diagnose noisy experimental data and maintain high test accuracy in an anti-attack test. This shows that the QSAE has non-linear mapping capability and robustness. [ABSTRACT FROM AUTHOR]

The lack of labeled training samples restricts the improvement of Hyperspectral Remote Sensing Image (HRSI) classification accuracy based on deep learning methods. In order to improve the HRSI classification accuracy when there are few training samples, a Lightweight 3D Dense Autoencoder Network (L3DDAN) is proposed. Structurally, the L3DDAN is designed as a stacked autoencoder which consists of an encoder and a decoder. The encoder is a hybrid combination of 3D convolutional operations and 3D dense block for extracting deep features from raw data. The decoder composed of 3D deconvolution operations is designed to reconstruct data. The L3DDAN is trained by unsupervised learning without labeled samples and supervised learning with a small number of labeled samples, successively. The network composed of the fine-tuned encoder and trained classifier is used for classification tasks. The extensive comparative experiments on three benchmark HRSI datasets demonstrate that the proposed framework with fewer trainable parameters can maintain superior performance to the other eight state-of-the-art algorithms when there are only a few training samples. The proposed L3DDAN can be applied to HRSI classification tasks, such as vegetation classification. Future work mainly focuses on training time reduction and applications on more real-world datasets. [ABSTRACT FROM AUTHOR]

Recently, Zero-Shot Learning (ZSL) has gained great attention due to its significant classification performance for novel unobserved classes. As seen and unseen classes are completely disjoint, the current ZSL methods inevitably suffer from the domain shift problem when transferring the knowledge between the observed and unseen classes. Additionally, most ZSL methods especially those targeting the semantic space may cause the hubness problem due to their use of nearest-neighbor classifiers in high-dimensional space. To tackle these issues, we propose a novel pathway termed Regularized Label Relaxation-based Stacked Autoencoder (RLRSA) to diminish the domain difference between seen and unseen classes by exploiting an effective label space, which has some notable advantages. First, the proposed method establishes the tight relations among the visual representation, semantic information and label space using via the stacked autoencoder, which is beneficial for avoiding the projection domain shift. Second, by incorporating a slack variable matrix into the label space, our RLRSA method has more freedom to fit the test samples whether they come from the observed or unseen classes, resulting in a very robust and discriminative projection. Third, we construct a manifold regularization based on a class compactness graph to further reduce the domain gap between the seen and unseen classes. Finally, the learned projection is utilized to predict the class label of the target sample, thus the hubness issue can be prevented. Extensive experiments conducted on benchmark datasets clearly show that our RLRSA method produces new state-of-the-art results under two standard ZSL settings. For example, the RLRSA obtains the highest average accuracy of 67.82% on five benchmark datasets under the pure ZSL setting. For the generalized ZSL task, the proposed RLRSA is still highly effective, e.g., it achieves the best H result of 58.9% on the AwA2 dataset. [ABSTRACT FROM AUTHOR]

The effectiveness of deep learning models depends on their architecture and topology. Thus, it is essential to determine the optimal depth of the network. In this paper, we propose a novel approach to learn the optimal depth of a stacked AutoEncoder, called Dynamic Depth for Stacked AutoEncoders (DDSAE). DDSAE learns in an unsupervised manner the depth of a stacked AutoEncoder while training the network model. Specifically, we propose a novel objective function, aside from the AutoEncoder's loss function to optimize the network depth: The optimization of the objective function determines the layers' relevance weights. Additionally, we propose an algorithm that iteratively prunes the irrelevant layers based on the learned relevance weights. The performance of DDSAE was assessed using benchmark and real datasets. [ABSTRACT FROM AUTHOR]

Intraspecific nest parasitism is a phenomenon that attracts the attention of biologists since it helps in saving the endangered species such as Slender Billed Gull. The problem comes from the fact that a parasite female lays its eggs in the nest of another female (host) of the same species which causes the abandon of the nest by the host. This behavior causes a significant reduction in future birds number and leads to the expansion of this specie. Thus, there has been an urgent necessity to clean the nest from parasitic eggs. So, our aim is to build an automatic parasitic egg identification system based on egg visual features information. Our system uses deep learning models which have proven their success for image classification. Indeed, our system conduct an egg image’s pre-processing phase followed by Fast Beta Wavelet Network (FBWN) to extract the most efficient descriptors (shape, texture, and color). Then, these features will be inputted to the Stacked AutoEncoder for egg classification. Our proposed system, has been evaluated on 91-egg dataset collected from 31 clutches of eggs in Sfax region, Tunisia. Our model has given a parasitic egg identification accuracy of 89.9% which has outperformed the state-of-the-art method and shows the efficiency and the robustness of our system.

The imaging spectrometer is limited by a short response time and narrow channel, resulting in a low signal-to-noise ratio of hyperspectral images. The accurate estimation of noise has a significant impact on some preprocessing and downstream tasks. The existing noise estimation methods for hyperspectral images are all focused on satellite and aviation data, and there is little research on hyperspectral images with high spatial resolution. For this type of image, This article proposes a noise estimation method based on a stacked autoencoder. Firstly, the image is divided into multiple uniform regions using the K-means algorithm, and then a stacked automatic encoder is set for each uniform region. Reconstruct the spectral signal on each pixel through the corresponding stacked automatic encoder. Calculate the residual between the reconstructed image and the original image to achieve signal-to-noise separation. Finally, the image is divided into a large number of subblocks, and the subblocks containing edges are removed. The remaining subblocks are used for noise estimation in this band. The applicability of some classic noise estimation methods was experimentally tested, and the effectiveness and stability of the proposed method were verified through simulation and real data experiments.

Hyperspectral band selection plays a key role for overcoming the curse of dimensionality in the classification of hyperspectral remote sensing images (HSIs). Recently, clustering-based band selection methods have demonstrated great potential to select informative and representative bands for hyperspectral classification tasks. However, most clustering-based methods perform clustering directly on the original high-dimensional data, which reduces their performance. To address this problem, a novel band selection method called global-local consistency constrained deep embedded clustering (GLC-DEC) is proposed in this paper. In GLC-DEC, to simultaneously learn the low-dimensional embedded representation and cluster assignments of all bands in an HSI, the stacked autoencoder is integrated with the K-means method. In addition, to reduce the adverse impact of a limited number of training samples available in HSIs, local and global consistency constraints are imposed on the embedded representation so that discriminatively consistent representation of all bands is learned. Specifically, local graph regularization and global graph regularization are introduced into the GLC-DEC model, by which the strong correlation between neighboring bands and the manifold structure of all bands are fully exploited. Based on the clustering results provided by GLC-DEC, a group of representative bands are selected by using the minimum noise method. Experimental results on two real datasets demonstrate that the proposed GLC-DEC outperformed several state-of-the-art methods.

Earthquake-induced landslides are ubiquitous on slopes in terrestrial environments, which can pose a serious threat to local communities and infrastructures. Data-driven landslide assessments play a crucial role in preventing future landslide occurrences and recurrences. We present a novel granular computing approach that assesses landslide risk by combining fuzzy information granulation and a stacked autoencoder algorithm. The stacked autoencoder is trained using an end-to-end learning strategy to obtain a central latent vector with reduced dimensionality. The multivariate landslide dataset was used as both the input and output to train the stacked autoencoder algorithm. Subsequently, in the central latent vector of the stacked autoencoder, the Fuzzy C-means clustering algorithm was applied to cluster the landslides into various groups with different risk levels, and the intervals for each group were computed using the granular computing approach. An empirical case study in Wenchuan County, Sichuan, China, was conducted. A comparative analysis with other state-of-the-art approaches including Density-based spatial clustering of applications with noise (DBSCAN), K-means clustering, and Principal Component Analysis (PCA), is provided and discussed. The experimental results demonstrate that the proposed approach using a stacked autoencoder integrated with fuzzy information granulation provides superior performance compared to those by other state-of-the-art approaches, and is capable of studying deep patterns in earthquake-induced landslide datasets and provides sufficient interpretation for field engineers.

Among the barriers to establishing effective human-machine interactions is the machines' inability to properly distinguish emotions from the human voice. The Speech Emotion Recognition (SER) systems have emerged to tackle this limitation. The accuracy of these systems depends on different factors such as the quantity and the types of emotions included in the database, feature extraction process including local and global features, feature selection method, and the type of classifier. This study presents a methodology for speech emotion recognition using an autoencoder neural network. It is shown that using a digit-level stacked autoencoder can be suitable for digit classification. The speech emotion recognition is done using the Persian emotional speech database (Persian ESD), which includes six emotional states: Happiness, Sadness, Fear, Disgust, Anger, and Neutral. Moreover, the popular, widely-used Berlin Emotional database (EMO-DB) is used to evaluate the effectiveness of the proposed approach. The experimental results show that the proposed method has significantly improved recognition accuracy. [ABSTRACT FROM AUTHOR]

The efficiency and cognitive limitations of manual sample labeling result in a large number of unlabeled training samples in practical applications. Making full use of both labeled and unlabeled samples is the key to solving the semi-supervised problem. However, as a supervised algorithm, the stacked autoencoder (SAE) only considers labeled samples and is difficult to apply to semi-supervised problems. Thus, by introducing the pseudo-labeling method into the SAE, a novel pseudo label-based semi-supervised stacked autoencoder (PL-SSAE) is proposed to address the semi-supervised classification tasks. The PL-SSAE first utilizes the unsupervised pre-training on all samples by the autoencoder (AE) to initialize the network parameters. Then, by the iterative fine-tuning of the network parameters based on the labeled samples, the unlabeled samples are identified, and their pseudo labels are generated. Finally, the pseudo-labeled samples are used to construct the regularization term and fine-tune the network parameters to complete the training of the PL-SSAE. Different from the traditional SAE, the PL-SSAE requires all samples in pre-training and the unlabeled samples with pseudo labels in fine-tuning to fully exploit the feature and category information of the unlabeled samples. Empirical evaluations on various benchmark datasets show that the semi-supervised performance of the PL-SSAE is more competitive than that of the SAE, sparse stacked autoencoder (SSAE), semi-supervised stacked autoencoder (Semi-SAE) and semi-supervised stacked autoencoder (Semi-SSAE). [ABSTRACT FROM AUTHOR]

An emergency risk assessment by collecting disaster-affected images via unmanned aerial vehicles is the current norm. Reasonable rescue planning and resource allocation depend on a quick and precise semantic interpretation of natural disaster images. However, the poor image quality produced by various technological and environmental factors and complex scenarios associated with disaster-affected regions makes the classification operation challenging. In order to get in-depth spatial features for decoding the intricate textures associated with catastrophe images, this study proposes an implementation of the CNN-based multibranch feature extraction technique. An advanced mixed-attention mechanism is exploited to extract the highly essential features. This mixed-attention mechanism effectively overcomes the flaws generated by traditional convolution by neglecting the global information and focusing on local key features. An SRGAN-based super-resolution method is utilized to acquire high-resolution images with rich spatial details to enhance the quality of aerial images. Besides, we experiment with several existing image classification algorithms, such as the ensemble model of pre-trained networks, the capsule network model, and the stacked autoencoder. Finally, we perform a comparative analysis between all the deployed models to obtain the best-performing classifier. Our proposed multibranch feature extraction with mixed-attention mechanism-based network performs more superiorly among the four models due to its ability to extract highly relevant features from disaster images. Generated super-resolution images effectively increase the classification performance. Our research findings and approaches accommodate quality resources for disaster image quality enhancement and classification activities. [ABSTRACT FROM AUTHOR]

The evolution of new trends in the automobile industry is creating more comfortable and convenient means of transportation. But still there exist manifold challenges in detecting legal and illegal drivers; therefore, evaluating the behaviors of drivers needs to be addressed. Taking these into consideration his paper proposes a novel cascade region convolutional neural network-based circulatory system optimization algorithm (CRCNN++ based CSO) to attain optimal multitask framework which includes behavior evaluation, identity authentication of drivers, vehicle diagnosis as well as estimating the fault of the vehicle. In this paper, two diverse naturalistic driving behavior public datasets namely HCRL and UAH drive datasets are collected and pre-processed via normalization as well as scaling process. The preprocessed feature is then extracted and the dimensions are minimized using the stacked autoencoder technique. The CRCNN++-based CSO is employed in determining to multitask which includes identity authentication, behavioral evaluation, vehicle diagnosis, and faults estimation is performed. Finally, the efficiency of the proposed CRCNN++-based CSO method is analyzed by evaluating various metrics namely receiver operating characteristic curve, accuracy, false positive rate, precision, Cohen Kappa score, true positive rate, and F1-Score. The comparative analysis is carried out for various existing techniques and the proposed approach. From the evaluation results, it is revealed that the proposed CRCNN++-based CSO approach delivers better performance in driver identification through driving style behavior. [ABSTRACT FROM AUTHOR]

Autism screening is crucial for the early diagnosis of developmental disorder. The combination of machine learning (ML) and deep learning (DL) approaches are applied to produce memory efficient and less complex deep learning models for the computer aided diagnosis (CAD) of autism screening. In the proposed work, two novel integrated activation functions such as Li-ReLU and S-RReLU are developed to aid in the classification of autistic subjects and typical controls (TC) with maximum accuracy. As functional magnetic resonance imaging (fMRI) data is noisy, it undergoes temporal and spatial pre-processing. The artifact free high dimensional fMRI data is exercised for the process of feature extraction and dimensionality reduction employing group principal component analysis (Group PCA) and group independent component analysis (Group ICA). The selected features are normalized using 0–1 normalization and converted to tensors. Stacked autoencoder (SAE) utilizes the fMRI tensor data for the classification of autism spectrum disorder (ASD) subjects and typical controls. The proposed work is implemented and tested on all datasets of ABIDE I database. The validation accuracy of CMU_a, KKI, UCLA_2, OLIN, Yale and NYU datasets are obtained as 100, 80, 71.43, 100, 85.71 and 93.33% using novel Li-ReLU activation function in the proposed system. With the help of new activation function called S-RReLU, the proposed system achieves validation accuracy of about 10, 100, 57.14, 100, 78.57 and 93.33% for CMU_a, KKI, UCLA_2, OLIN, Yale and NYU datasets. Thus, the proposed method outperforms all other existing state-of-the-art works in terms of accuracy. [ABSTRACT FROM AUTHOR]

Modern industrial processes often exhibit large-scale and nonlinear characteristics. Incipient fault detection for industrial processes is a big challenge because of the faint fault signature. To improve the performance of incipient fault detection for large-scale nonlinear industrial processes, a decentralized adaptively weighted stacked autoencoder (DAWSAE) -based fault detection method is proposed. First, the industrial process is divided into several sub-blocks and local adaptively weighted stacked autoencoder (AWSAE) is established for each sub-block to mine local information and obtain local adaptively weighted feature vectors and residual vectors. Second, the global AWSAE is established for the whole process to mine global information and obtain global adaptively weighted feature vectors and residual vectors. Finally, local statistics and global statistics are constructed based on local and global adaptively weighted feature vectors and residual vectors to detect the sub-blocks and the whole process, respectively. The advantages of proposed method are verified by a numerical example and Tennessee Eastman process (TEP). • A decentralized framework is proposed to fully consider the local and global information. • An adaptively weighted SAE model is designed to enhance the feature of incipient faults. • Local and global statistics are constructed respectively to perform preliminary fault location. [ABSTRACT FROM AUTHOR]

It is crucial in industrial processes to consider key variables to ensure safe operation and high product quality. Moreover, these variables are difficult to obtain using traditional measurement methods; hence, it makes sense to develop soft sensor regression models to process the variable prediction. However, there are numerous variables integrating noisy and redundant information in complex industrial processes. Using such variables in traditional regression models may result in reducing the model's efficiency and performance. Thus, this paper proposes a multi‐layer feature ensemble soft sensor regression method using a stacked auto‐encoder (SAE) and vine copula (ESAE–VCR) to address these problems. To do so, the number of neurons in the hidden layer of the SAE is determined by the principal component analysis (PCA). The multi‐layer features of the process variables are extracted using a stacked AE, and the regression models are established for each feature layer. A linear regression ensemble method is used to combine the regression models with the multi‐layer features to obtain the final predictive model that will estimate the values of the key variables. The effectiveness and practicality of the ESAE–VCR are validated by comparing them with several common soft measurement methods in two examples. In the numerical example, the ESAE–VCR yields an accuracy of prediction (R2) of 0.9898 and a root mean square error (RMSE) of 0.1804. In the industrial example, the ESAE–VCR yields an R2 of 0.9908 and an RMSE of 0.1205. [ABSTRACT FROM AUTHOR]

The goal of the paper is to present a Stacked autoencoder approach for enhancing Intrusion Detection Systems (IDSs) in Mobile Ad-Hoc Networks (MANETs). The paper proposes a Stacked autoencoder-based approach for MANET (Stacked AE-IDS) to reduce correlation and model relevant features with high-level representation. This method reproduces the input with a reduced correlation, and the output of the autoencoder is used as the input of the Deep Neural Network (DNN) classifier (DNN-IDS). The proposed Deep Learning-based IDS focuses on Denial of Services (DoS) attacks within labeled datasets, which are available for intrusion detection, and employs the most potential attacks that impact routing services in Mobile Networks. The proposed Stacked AE-IDS method enhances the effectiveness of IDSs in detecting attacks in MANETs by reducing the correlation and modeling high-level representations of relevant features. The focus on DoS attacks and their impact on routing services in Mobile Networks makes the proposed approach particularly relevant for MANET security. The proposed Stacked AE-IDS approach has potential applications in enhancing the security of MANETs by improving the effectiveness of IDSs. This approach can be used to detect different types of attacks, particularly DoS attacks, and their impact on routing services in Mobile Networks. [ABSTRACT FROM AUTHOR]