Neural Computing and Applications最新文献

Electronic newspapers are one of the most common sources of Modern Standard Arabic. Existing datasets of Arabic news articles typically provide a title, body, and single label. Ignoring important features, like the article author, image, tags, and publication date, can degrade the efficacy of classification models. In this paper, we propose the Arabic multi-purpose integral news articles (AMINA) dataset. AMINA is a large-scale Arabic news corpus with over 1,850,000 articles collected from 9 Arabic newspapers from different countries. It includes all the article features: title, tags, publication date and time, location, author, article image and its caption, and the number of visits. To test the efficacy of the proposed dataset, three tasks were developed and validated: article textual content (classification and generation) and article image classification. For content classification, we experimented the performance of several state-of-the-art Arabic NLP models including AraBERT and CAMeL-BERT, etc. For content generation, the reformer architecture is adopted as a character text generation model. For image classification applied on Al-Sharq and Youm7 news portals, we have compared the performance of 10 pre-trained models including ConvNeXt, MaxViT, ResNet18, etc. The overall study verifies the significance and contribution of our newly introduced Arabic articles dataset. The AMINA dataset has been released at https://huggingface.co/datasets/MohamedZayton/AMINA.

Stochastic optimizers play a crucial role in the successful training of deep neural network models. To achieve optimal model performance, designers must carefully select both model and optimizer hyperparameters. However, this process is frequently demanding in terms of computational resources and processing time. While it is a well-established practice to tune the entire set of optimizer hyperparameters for peak performance, there is still a lack of clarity regarding the individual influence of hyperparameters mislabeled as “low priority”, including the safeguard factor (epsilon) and decay rate (beta), in leading adaptive stochastic optimizers like the Adam optimizer. In this manuscript, we introduce a new framework based on the empirical probability density function of the loss’ gradient magnitude, termed as the “gradient magnitude histogram”, for a thorough analysis of adaptive stochastic optimizers and the safeguard hyperparameter (epsilon). This framework reveals and justifies valuable relationships and dependencies among hyperparameters in connection to optimal performance across diverse tasks, such as classification, language modeling and machine translation. Furthermore, we propose a novel algorithm using gradient magnitude histograms to automatically estimate a refined and accurate search space for the optimal safeguard hyperparameter (epsilon), surpassing the conventional trial-and-error methodology by establishing a worst-case search space that is two times narrower.

This paper explores the importance of the electric eel foraging optimization (EEFO) algorithm in addressing feature selection (FS) problems, with the aim of ameliorating the practical benefit of FS in real-world applications. The use of EEFO to solve FS problems props our goal of providing clean and useful datasets that provide robust effectiveness for use in classification and clustering tasks. High-dimensional feature selection problems (HFSPs) are more common nowadays yet intricate where they contain a large number of features. Hence, the vast number of features in them should be carefully selected in order to determine the optimal subset of features. As the basic EEFO algorithm experiences premature convergence, there is a need to enhance its global and local search capabilities when applied in the field of FS. In order to tackle such issues, a binary augmented EEFO (BAEEFO) algorithm was developed and proposed for HFSPs. The following strategies were integrated into the mathematical model of the original EEFO algorithm to create BAEEFO: (1) resting behavior with nonlinear coefficient; (2) weight coefficient and confidence effect in the hunting process; (3) spiral search strategy; and (4) Gaussian mutation and random perturbations when the algorithm update is stagnant. Experimental findings confirm the effectiveness of the proposed BAEEFO method on 23 HFSPs gathered from the UCI repository, recording up to a 10% accuracy increment over the basic BEEFO algorithm. In most test cases, BAEEFO outperformed its competitors in classification accuracy rates and outperformed BEEFO in 90% of the datasets used. Thereby, BAEEFO has demonstrated strong competitiveness in terms of fitness scores and classification accuracy. When compared to its competitors, BAEEFO produced superior reduction rates with the fewest number of features selected. The findings in this research underscore the critical need for FS to combat the curse of dimensionality concerns and find highly useful features in data mining applications such as classification. The use of a new meta-heuristic algorithm incorporated with efficient search strategies in solving HFSPs represents a step forward in using this algorithm to solve other practical real-world problems in a variety of domains.

In order to maintain sustainable agriculture, it is vital to monitor plant health. Since all species of plants are prone to characteristic diseases, it necessitates regular surveillance to search for any symptoms, which is utterly challenging and time-consuming. Besides, farmers may struggle to identify the type of plant disease and its potential symptoms. Hence, the interest in research like image-based computer-aided automated plant leaf disease detection by analyzing the early symptoms has increased enormously. However, limitations in the plant leaf image database, for instance, unfitting backgrounds, blurry images, and so on, sometimes cause underprivileged feature extraction, misclassification, and overfitting issues in existing models. As a result, we have proposed a real-time plant leaf disease detection architecture incorporating proposed PLD-Det model, which is based on improved YOLOv7 with the intention of assisting farmers while reducing the issues in existing models. The architecture has been trained on the widely used PlantVillage dataset, which resulted in an accuracy of 98.53%. Furthermore, SHapley Additive exPlanations (SHAP) values have been analyzed as a unified measure of feature significance. According to the experimental findings, the proposed PLD-Det model, which is an improved YOLOv7 architecture, outperformed the original YOLOv7 model in test accuracy by approximately 4%.

Automatic age and gender estimation provides an important information to analyze real-world applications such as human–machine interaction, system access, activity recognition, and consumer profile detection. While it is easy to estimate a person’s gender from human facial images, estimating their age is difficult. In such previous challenging studies, traditional convolutional neural network (CNN) methods have been used for age and gender estimation. With the development of deep convolutional neural network (DCNN) architectures, more successful results have been obtained than traditional CNN methods. In this study, two state-of-the-art DCNN models have been developed in the field of artificial intelligence (AI) to make age and gender estimation on an imbalanced dataset of human face images. Firstly, a new model called fast description network (FINet) was developed, which has a parametrically changeable structure. Secondly, the number of parameters has been reduced by using the layer reduction approach in InceptionV3 and NASNetLarge DCNN model structures, and a second model named inception Nasnet fast identify network (INFINet) was developed by concatenating these models and the FINet model as a triple. FINet and INFINet models developed for age and gender estimation were compared with many other state-of-the-art DCNN models in AI. The most successful accuracy results in terms of both age and gender were obtained with the INFINet model (age: 61.22%, gender: 80.95% in the FG-NET dataset, age: 72.00%, gender: 90.50% in the UTKFace dataset). The results obtained in age and gender estimation with the INFINet model are much more effective than other recent state-of-the-art works. In addition, the FINet model, which has a much smaller number of parameters than the compared models, showed a classification performance that can compete with state-of-the-art methods for age and gender estimation.

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that mainly affects memory and other cognitive functions, such as thinking, reasoning, and the ability to carry out daily activities. It is considered the most common form of dementia in older adults, but it can appear as early as the age of 25. Although the disease has no cure, treatment can be more effective if diagnosed early. In diagnosing AD, changes in the brain’s morphology are identified macroscopically, which is why deep learning models, such as convolutional neural networks (CNN) or vision transformers (ViT), excel in this task. We followed the Systematic Literature Review process, applying stages of the review protocol from it, which aims to detect the need for a review. Then, search equations were formulated and executed in several literature databases. Relevant publications were scanned and used to extract evidence to answer research questions. Several CNN and ViT approaches have already been tested on problems related to brain image analysis for disease detection. A total of 722 articles were found in the selected databases. Still, a series of filters were performed to decrease the number to 44 articles, focusing specifically on brain image analysis with CNN and ViT methods. Deep learning methods are effective for disease diagnosis, and the surge in research activity underscores its importance. However, the lack of access to repositories may introduce bias into the information. Full access demonstrates transparency and facilitates collaborative work in research.

Credit card fraud results in staggering financial losses amounting to billions of dollars annually, impacting both merchants and consumers. In light of the escalating prevalence of digital crime and online fraud, it is important for organizations to implement robust and advanced technology to efficiently detect fraud and mitigate the issue. Contemporary solutions heavily rely on classical machine learning (ML) and deep learning (DL) methods to handle such tasks. While these methods have been effective in many aspects of fraud detection, they may not always be sufficient for credit card fraud detection as they aren’t adaptable to detect complex relationships when it comes to transactions. Fraudsters, for example, might set up many coordinated accounts to avoid triggering limitations on individual accounts. In the context of fraud detection, the ability of Graph Neural Networks (GNN’s) to aggregate information contained within the local neighbourhood of a transaction enables them to identify larger patterns that may be missed by just looking at a single transaction. In this research, we conduct a thorough analysis to evaluate the effectiveness of GNNs in improving fraud detection over classical ML and DL methods. We first build an heterogeneous graph architecture with the source, transaction, and destination as our nodes. Next, we leverage Relational Graph Convolutional Network (RGCN) to learn the representations of nodes in our graph and perform node classification on the transaction node. Our experimental results demonstrate that GNN’s outperform classical ML and DL methods.

Click-through rate (CTR) prediction aims to estimate the probability of a user clicking on a particular item, making it one of the core tasks in various recommendation platforms. In such systems, user behavior data are crucial for capturing user interests, which has garnered significant attention from both academia and industry, leading to the development of various user behavior modeling methods. However, existing models still face unresolved issues, as they fail to capture the complex diversity of user interests at the semantic level, refine user interests effectively, and uncover users’ potential interests. To address these challenges, we propose a novel model called knowledge-enhanced Interest segment division attention network (KISDAN), which can effectively and comprehensively model user interests. Specifically, to leverage the semantic information within user behavior sequences, we employ the structure of a knowledge graph to divide user behavior sequence into multiple interest segments. To provide a comprehensive representation of user interests, we further categorize user interests into strong and weak interests. By leveraging both the knowledge graph and the item co-occurrence graph, we explore users’ potential interests from two perspectives. This methodology allows KISDAN to better understand the diversity of user interests. Finally, we extensively evaluate KISDAN on three benchmark datasets, and the experimental results consistently demonstrate that the KISDAN model outperforms state-of-the-art models across various evaluation metrics, which validates the effectiveness and superiority of KISDAN.

Alzheimer's disease is a progressive form of dementia. Dementia is a broad term for conditions that impair memory, thinking, and behaviour. Brain traumas or disorders can cause dementia. It is estimated that 60–80% of dementia cases around the world are caused by Alzheimer’s disease, an incurable neurodegenerative disorder. Although Alzheimer's disease research has increased in recent years, early diagnosis is challenging due to the complicated brain structure and functions associated with this disease. It is difficult for doctors to identify Alzheimer's disease in its early stages as there are still no biomarkers to be precise in early detection. In the area of medical imaging, deep learning is becoming increasingly popular and successful. There is no single best approach for the detection of Alzheimer's disease. In comparison with conventional machine learning methods, the deep learning models detect Alzheimer's disease more precisely and effectively. In this review paper, various machine learning-based techniques utilized for the classification of Alzheimer's disease through different imaging modalities are discussed. In addition, a comprehensive and detailed analysis of the various image processing procedures along with corresponding classification performance and feature extraction techniques have been meticulously compiled and presented. The investigation of computer-aided image analysis has demonstrated significant potential in the early detection of cognitive changes in individuals experiencing mild cognitive impairment. Machine learning can provide valuable insights into the cognitive status of patients, enabling healthcare professionals to intervene and provide timely treatment. This review may lead to a reliable method for recognizing and predicting Alzheimer's disease.

Prompt-based portrait image style transfer aims at translating an input content image to a desired style described by text without a style image. In many practical situations, users may not only attend to the entire portrait image but also the local parts (e.g., eyes, lips, and hair). To address such applications, we propose a new framework that enables style transfer on specific regions described by a text description of the desired style. Specifically, we incorporate semantic segmentation to identify the intended area without requiring edit masks from the user while utilizing a pre-trained CLIP-based model for stylizing. Besides, we propose a text-to-patch matching loss by randomly dividing the stylized image into smaller patches to ensure the consistent quality of the result. To comprehensively evaluate the proposed method, we use several metrics, such as FID, SSIM, and PSNR on a dataset consisting of portraits from the CelebAMask-HQ dataset and style descriptions of other related works. Extensive experimental results demonstrate that our framework outperforms other state-of-the-art methods in terms of both stylization quality and inference time.