Floods remain among the most destructive natural hazards, presenting severe threats to human life, infrastructure, and ecological systems. As extreme weather events become increasingly frequent due to climate change, the development of accurate and interpretable flood prediction models is more critical than ever. This study presents a deep tabular learning framework for flood risk prediction using TabNet, a neural architecture optimized for structured tabular data. The model is trained on a real-world dataset sourced from Kaggle, consisting of 50,000 records with 21 environmental and socio-economic indicators, including features like monsoon intensity, topography, deforestation, urbanization, and drainage systems. The dataset contains no missing values or categorical variables, enabling direct application of regression models. After preprocessing and standardization, the TabNet model was trained with optimized hyperparameters using a five-step decision process and sparse attention mechanisms to identify feature importance during learning. The model achieved outstanding predictive performance, with a Mean Absolute Error (MAE) of 0.0069, Mean Squared Error (MSE) of 0.000094, Root Mean Squared Error (RMSE) of 0.0097, Mean Absolute Percentage Error (MAPE) of 1.36

The problem of academic performance prediction has beco- me one of the most pressing educational data mining studies as it can facilitate early detection of at-risk students and enable the provision of appropriate academic support as soon as possible. Attendance, study habits, and sleep patterns are some of the elements of behavior that are very crucial in determining the practice of learning but the combined effects have not been well investigated in predictive modelling. In this paper, the researchers offer an artificial intelligence-based model of academic performance prediction through the analysis of attendance rates, indicators of study behavior, and sleep-related features. Several machine learning models are tested on artificial learning going through an actual educational dataset. Empirical evidence shows that the ensemble-based methods have better predictive power that non ensemble-based classifiers. The suggested framework is an economical and evidence-based approach that allows educators to plan academically forward and implement individualized learning plans.

India’s urban traffic system continues to suffer from severe congestion, inefficient signal timing, and increasing vehicle volumes, resulting in prolonged travel times, higher emissions, and frequent rule violations. To address these challenges, this study presents Smart Traffic Eye, a real-time traffic control system powered by deep learning. Leveraging the YOLO object detection algorithm for accurate and fast vehicle identification, the system integrates with the SUMO (Simulation of Urban Mobility) platform to simulate and manage dynamic traffic signal control. The model was trained on the MIO-TCD dataset, which includes diverse vehicle categories and environmental conditions representative of Indian roads. By analyzing real-time traffic density and adjusting signal timing accordingly, the system prioritizes emergency vehicles and reduces congestion. Simulation results demonstrate a clear improvement in traffic throughput, reduced average waiting times, and better adaptability to fluctuating traffic loads. The proposed framework not only offers a scalable and cost-effective solution but also aligns with India’s smart city goals by paving the way for AI-driven, responsive urban mobility systems.

The aim of this study is to conduct a systematic literature review using bibliometric analysis techniques on the research articles published between 2000 and 2022 on the usage and adoption of mobile payment. Since mobile payments act as a strong driver of achieving sustainable development goals and enable users to conduct transactions anytime and anywhere, there has been an increase in scientific research on this topic. Moreover, mobile payments have significantly contributed in achieving SDG 13, 1, 8 and 10, which are related to climate action, no poverty, decent work, and reduce inequality. Because of the large number of publications available on this topic, it is crucial to identify the most prominent authors, papers, journals, and organizations. Thus, data for the purpose of bibliometric analysis was gathered from the Scopus database, and 456 articles were found suitable. The data was analyzed using VOSviewer software, and tables, graphs, and network visualizations were used to demonstrate the trends in publishing years and list the most prominent authors, papers, journals, and organizations. Moreover, this study adds to the existing body of knowledge by offering directions for further research. The framework gives researchers the chance to investigate the study streams in their upcoming studies.

Email spam detection is an essential task in securing digital communication systems. This study, introduce SpamXFormer, a novel deep learning framework designed to enhance email classification through the integration of semantic, structural, and contextual cues. The experiments utilize a publicly available Kaggle dataset consisting of 83,446 email records, categorized as spam (1) or ham (0), with no missing entries. Each record contains raw email text, which varies widely in content, structure, and length. SpamXFormer combines DistilBERT-based semantic embeddings, a 1D convolutional layer for capturing local n-gram patterns, and a self-attention mechanism to highlight spam-indicative tokens. The architecture further incorporates auxiliary structural features such as normalized email length and word entropy to improve robustness. This hybrid design enables SpamXFormer to model both long-range dependencies and localized spam signals. The model achieved an impressive accuracy of 96.75

Mental sickness is of great concern to the world these days! On the other hand, the provision of psychological services is mainly restricted not only because society usually stigmatizes it, but also due to the developments in the field, lack of skilled professionals, and above all, misunderstandings. It is the chatbots based on the AI technology that have the potential for scalability and accessibility which may be the answer to this problem. In this article, the new chatbot Soul Sync will be introduced and defined as a chatbot using artificial intelligence for enhancing the quality of mental healthcare, the increase of user engagement, the decrease of TMs, and the creation of golden digital therapeutic relationship. The current research followed a mixed methods design consisting of a comprehensive literature review, user-centered design methods, and survey plus interview data collection by empirical means. It should be noted that the study that is at hand examined the ease with which AI chatbots could be employed literally and their efficacy in dealing with young adults where media users. The authors not only show that the use of AI chatbots can reduce mental health problems to a considerable extent and produce cognitive-behavioral therapy performance but also be an instrumental factor in health. First and foremost, the investigation discovered that AI represented a major lift in mental health, and one that maybe assigned to treating of special psychological disorders, as well as the changing of people’s habits. Besides, the paper argues that the efficient use of AI chatbots in this domain may still be hindered by several challenges such as designing a front end interface that is user-friendly, algorithmic bias, privacy concerns and setting it up in the current health care infrastructure.

High-dimensional gene expression datasets with limited samples pose a persistent challenge for machine learning in cancer genomics. To address this, we propose BioFuseNet, a novel deep learning architecture designed to classify leukemia subtypes with high accuracy by combining autoencoder-based representation learning and gene-set-aware feature fusion. The model operates on a carefully curated microarray dataset from the CuMiDa (Curated Microarray Database), consisting of 64 samples, 5 leukemia classes, and over 22,000 gene expression features. Leveraging unsupervised gene clustering, BioFuseNet divides the input features into distinct subsets, each processed through a dedicated autoencoder branch that learns compact and denoised representations. These branch-wise embeddings are then fused and passed through a fully connected classifier trained jointly using both supervised cross-entropy loss and unsupervised reconstruction loss. Experimental evaluation shows that BioFuseNet achieves a classification accuracy of 90

The cardiovascular diseases are among the major causes of death in the global population, and therefore, they require efficient and prompt diagnostic systems to aid the clinical decision-making process. Recent developments in machine learning have shown high potential predicting cardiac illness by means of multivariate clinical data. The paper describes an early heart disease detection framework using the power of machine learning with the use of structured clinical indicators, such as demographic characteristics, physiological, and the results of diagnostic tests. Several supervised learning models are tested, and the performance is measured with the help of conventional classification measures. According to the results of the experiment, ensemble and tree-based ideas can be used to obtain better predictive accuracy, than traditional classifiers. The proposed framework provides a valid, evidence-based tool that could be used to help the clinicians to perform proactive cardiovascular risk assessment and early intervention.

The global pharmaceutical industry faces a critical challenge with the proliferation of counterfeit medicines, posing severe health risks and economic losses. This research presents an AI-powered counterfeit medicine detection system that integrates YOLO (You Only Look Once), Convolutional Neural Networks (CNNs), and EasyOCR for comprehensive authentication. The system leverages YOLO for real-time object detection to analyze packaging anomalies, CNNs for feature-based classification, and OCR for text verification against authentic drug databases. In addition, adaptive weight adjustment is used to optimize the decision-making of the ensemble model, ensuring higher precision and robustness. The proposed methodology is tested using a publicly available Roboflow dataset, with the system optimized for high-performance GPUs (A100, L4) to handle computationally intensive tasks. Extensive evaluations are conducted using standard performance metrics, including mAP, Precision, Recall, F1-Score for YOLO; Accuracy, Confusion Matrix for CNN; Text Similarity Score for OCR; and a final weighted confidence score for the ensemble model. The results demonstrate significant improvements in the detection of counterfeit medicines, with high accuracy and detection reliability. This research contributes to the development of an efficient, scalable, and secure AI-driven pharmaceutical authentication framework.

Personalized Federated Learning (PFL) is an advanced machine learning approach that customizes global models to individual user data while maintaining privacy through decentralized training. This study presents a bibliometric analysis of PFL research using data sourced from the Scopus database. Three bibliometric tools—Biblioshiny, VOSviewer, and CiteSpace—were employed to examine publication trends, author impact, thematic evolution, and scholarly networks. The analysis highlights a sharp rise in annual scientific production, indicating rapid growth in the field since 2020. Most relevant authors and sources, including Chen Zihan and IEEE Transactions on Mobile Computing, demonstrate concentrated influence in shaping the domain. Thematic mapping reveals well-developed core themes such as personalization and data privacy, alongside emerging topics like federated reinforcement learning and hypernetworks. Network visualizations of co-cited authors and keyword co-occurrences provide insights into intellectual structure and evolving research priorities. Identified research gaps and underexplored themes suggest practical implications for expanding PFL into diverse domains such as IoT, healthcare, and privacy-preserving edge intelligence.

Biomedical image segmentation demands exceptional precision, surpassing that typically required for natural image processing. The U-Net neural network, renowned in biomedical applications, integrates multiple auto-encoder elements to generate comprehensive hypotheses for extensive image segmentation. This work introduces U-Net++ architecture, characterised by dense skip connections and a cascade of encoder–decoder modules tailored for tasks such as chest CT scan diagnosis, nuclei and polyp segmentation, and MRI-based heart analysis. Comparative assessments reveal that U-Net++ achieves superior Intersection over Union (IOU) metrics compared to conventional U-Net architecture. Efficient delineation of brain tumour extent is a critical clinical need. With the addition of selective kernels and squeeze and excitation residuals to the U-Net framework, the automated method presented in this study produced an impressive mean dice score of 0.992 for Late Gadolinium Enhancement (LGE) segmentation in left ventricle (LV), right ventricle (RV), and left ventricular myocardium (LVM). This research introduces the NN U-Net network paradigm to the medical domain by recognising the dependencies inherent in preprocessing, training, and inference with the original U-Net model. NN U-Net exhibits adaptability and robustness, yielding superior dice scores across all classes. This study explores U-Net networks and their various architectural adaptations, documenting contributions published in biomedical image analysis.

In the fast-evolving world of social media, user engagement is key to platform success. This study presents a novel approach to enhancing engagement through advanced classification algorithms for personalized content delivery, moving beyond generic strategies. The framework analyzes user behavior to provide tailored recommendations, adapting to changing interests and improving the overall experience. The classification algorithms effectively identify user preferences, resulting in more relevant content and higher interaction rates. The implementation and impact of these algorithms demonstrate that personalized engagement boosts content discoverability and strengthens user-platform relationships. Additionally, this article introduces a technology for classifying Facebook users using Particle Swarm Optimization (PSO). As social media evolves, this research aims to refine engagement strategies, highlighting the need for personalized content delivery to create a user-centric experience.

The shopping pattern of common people is undergoing a significant transition in the context of the emerging digital era. This is very evident among the young generation. There are many factors that cause online purchase decisions to buy products but emotions and preferences play a significant role in this. During online shopping, the consumer is engaged in a cautious evaluation of products and this may be attributed to the rapid popularization of internet technologies. They conduct price comparisons and analysis before making an online purchase decision. Currently, e-commerce is in the process of accommodating new technology-based marketing techniques to influence the buying decisions of young customers on online platforms. One of the latest technologies becoming popular in business activities is Robotic Process Automation (RPA). In the field of e-commerce, RPA plays an important role in automating tasks such as targeted advertising and customer service. This paper focuses on the aspects of RPA techniques majorly employed by e-commerce sites. A survey is conducted among young online shoppers to understand how these techniques affect their online impulsive purchase decisions. This paper also explores the improved customer experience that is achieved as a result of RPA technologies presented to online customers. The paper ends with future research directions and ethical considerations toward RPA integration in online stores.

Everyday new technologies are emerging to ensure that access to sensitive information is limited to only those with the authorized credentials. Recently, gait biometric recognition technique have attracted much attention and achieved remarkable success in many computer vision based security and surveillance techniques. This paper proposes an improved human recognition solution for Gait recognition that aims to ensure security access to only individuals with a registered Gait in the system. Proposed novel human Gait framework is based on the KNN-SIFT (K-Nearest Neighbours - Scale Invariant Feature Transformation) classifier algorithm that can be used as biometric security systems. The proposed tracking system is developed to recognize human based on built-in detection, training, testing and recognition options. The proposed system is able to identify human at a distance even with low resolution. The experimental results show that the proposed KNN-SIFT achieve 96.5% accuracy score and is comparable to state-of-the-art algorithms.

The rising prevalence of hematological malignancies, such as leukemia, in India underscores the critical need for enhanced diagnostic methods that enable early detection and treatment. This study embarks on a thorough examination of existing deep learning (DL) 1 methodologies applied to the diagnosis of hematological cancers, highlighting pivotal advancements and identifying prevailing gaps in current approaches. Without relying on actual datasets, our research synthesizes findings from extensive literature to propose a robust theoretical framework and a comprehensive mathematical model designed to enhance diagnostic accuracy. The proposed framework leverages advanced machine learning techniques, including enhanced Generative Adversarial Networks (GANs) 2 and Convolutional Neural Networks (CNNs) via sophisticated transfer learning processes. We introduce novel segmentation and classification algorithms that address specific challenges such as overlapping nuclei and morphological heterogeneity. The integration of Explainable AI (XAI) and principles of federated learning in our model underscores our commitment to maintaining transparency and safeguarding data privacy in clinical applications. By theoretical alignment and mathematical rigor, our proposed model aims to set a new benchmark in the diagnostic procedures of hematological malignancies, offering a scalable and adaptable solution that can be empirically validated in future research.

The study piece explores the rising suicide rates worldwide trend with a specific focus on the Indian context. By using AI algorithms and data mining approaches, it seeks to predict the root reasons for suicide. The intention is to enable specialists to take preventative action to avert such tragedies by pinpointing the critical elements that contribute to suicide behavior. An innovative LSTM-Attention-CNN model for analyzing internet entertainment content for indications of suicidal thoughts is presented in this study. The suggested model outperforms the current benchmark models with an astounding accuracy of 90.30