Browse Items (3095 total)
Sort by:
-
Integrating Explainability and Fairness in Credit Risk Prediction: A Hybrid Approach Using Tabnet, LightGBM, SHAP, and Counterfactual Explanations on the FICO Dataset
Transparent and fair credit risk assessment is essential for responsible lending in modern financial systems. This paper presents an interpretable and ethically grounded machine learning framework for loan default prediction using the FICO Explainability Challenge dataset. We combine LightGBM, a high-performing gradient boosting model for tabular data, with TabNet, a deep learning architecture that provides intrinsic interpretability through attentive feature selection. To enhance transparency, SHapley Additive exPlanations (SHAP) are employed for global and local feature attribution, while counterfactual explanations generated using the DiCE framework offer actionable recourse. Fairness is evaluated and mitigated using IBM's AI Fairness 360 toolkit. Experimental results demonstrate that the proposed hybrid approach achieves strong predictive performance while ensuring interpretability and fairness, making it suitable for trustworthy and regulation-compliant credit risk modeling. 2026 IEEE. -
Deep Q-Learning for Autonomous Vehicle Navigation in Smart Mobility
The proposed system leverages Deep Q-Learning to enhance autonomous vehicle navigation in smart mobility environments. By integrating reinforcement learning with deep neural networks, the system enables vehicles to make real-time decisions while adapting to dynamic traffic conditions. The framework employs a reward-based learning mechanism to optimize path selection, collision avoidance, and efficient maneuvering in complex urban scenarios. To improve decisionmaking accuracy, the proposed approach incorporates an experience replay mechanism, preventing overfitting and ensuring stable learning. Additionally, a target network is utilized to enhance training convergence, allowing the model to generalize effectively across varying road conditions. The system is further optimized through adaptive explorationexploitation strategies, enabling vehicles to balance learning new routes while prioritizing safe and efficient navigation. The proposed methodology demonstrates significant improvements in autonomous mobility, offering a scalable and robust solution for next-generation smart transportation systems. 2025 IEEE. -
Automated Leaf Disease Detection using a Hybrid CNN-BiLSTM Model for Smart Agriculture
The mitigation of crop losses and the sustainability of agriculture rely on the prompt identification of foliar diseases. In large-scale agriculture, conventional identification methods such as expert eye inspections are inefficient, susceptible to errors, and labour-intensive. A growing number of individuals are seeking automated methods to monitor plant health, given that the majority of Indians are employed in agriculture. This study presents a hybrid DL strategy for leaf disease detection, encompassing preprocessing, segmentation, feature extraction, and model training. Initially, images are processed to enhance their quality and uniformity. The impacted regions of the leaf are subsequently categorised by K-Means clustering. The classification accuracy is improved by utilising several feature extraction methods. The proposed model, CNBiLS, integrates bidirectional LSTM layers with convolutional layers to leverage the spatial and sequential information in image data. When evaluated against contemporary state-of-the-art models, CNBiLS exhibited superior performance, achieving an exceptional 99.84% classification accuracy. This result underscores the model's accuracy in identifying various leaf diseases. Ultimately, CNBiLS offers a precise, scalable, and robust automated system for detecting leaf diseases, equipping farmers with timely information to manage illnesses effectively, so enhancing both the quality and yield of their crops. 2025 IEEE. -
Hybrid Bidirectional GRU Approach for Crop Yield Prediction and Climate Change Impact Assessment in Agriculture
The impacts of climate change induced by humans will be felt most acutely by the agriculture sector due to its extreme dependence on weather. To ensure a steady supply of food, it is necessary to study and anticipate the effects of climate change on agricultural output. The impact of climate change on agricultural yield predictions is examined in this study using a novel methodology. In the proposed model, preprocessing, feature extraction, and training are the main processes. Data pretreatment guarantees quality by cleaning and normalising the data, while the PCC is utilised for feature selection. The model utilises AM and BiGRU for usage with large datasets. Using word vectors, the word embedding layer improves contextual awareness. Experiment findings show that the model is accurate to within 98.31% and can withstand a wide range of climate conditions. Current state-of-the-art methods are vastly outperformed by it, with performance measures like as R2 = 0.921%, MAE = 0.127%, and RMSE = 0.158%. These findings show that agricultural strategists and lawmakers can use AM-BiGRU to assess the effects of climate change and build a more resilient food system. 2025 IEEE. -
Predictive Modeling of Substance Abuse Risks using Big Data Analytics and Social Media Mining
The worldwide increase in substance abuse among teenagers and young adults has become serious concern in recent times. One way this pattern has developed is through the evolution of social media. Social media has transformed people's attitudes towards certain behaviors and has encouraged risky behavior to the point of actually causing addiction by exposing them to drug-related material. Despite the existence of preventative measures, such as education programs in schools, many children and youth have not had adequate access to educational interventions or evidence-based measures due to barriers created by geography, economic circumstances, and social factors, particularly in less developed countries. The research proposed is focusing on addressing this gap using a big data approach. This research employs a unique analytical framework that integrates multiple large data sets from a variety of sources to better identify and assess the effectiveness of interventions. This model employs an analytical approach that uses statistical learning techniques and predictive analytics to identify historical patterns and anticipate future trends, and assess the effectiveness of various interventions conducted in different countries. The results of the analysis suggest that this big data approach will provide decision-makers with clearly documented evidence related to various risk-taking behaviors as they relate to available prevention interventions, and will assist decision-makers in developing targeted prevention intervention strategies. This study demonstrates the revolutionary aspect behind the application of computational intelligence in preventing substance abuse and informing evidence-based community health interventions. 2025 IEEE. -
Digital Twins for predictive maintenance of Production and Machines: A Comprehensive Review
Digital Twin (DT) technology has matured from concept to practice across factories and critical assets, enabling new capabilities in condition-based and predictive maintenance, resilient production planning, and life-cycle decision-making. This review synthesizes current knowledge on DT usage for production and predictive machine maintenance, with concise notes on structural maintenance where SHM (structural health monitoring) increasingly adopts twin concepts. We first consolidate enabling architectures (standards, ontologies, F?FMI-based co-simulation, and hybrid modelling) and then critically survey applications spanning CNC cutting tools, bearings and gearboxes, robotic cells, and production lines. We highlight evidence that hybrid (physics + data-driven) twins reduce remaining useful life (RUL) prediction error compared to single-strategy approaches, improve energy-aware scheduling, and shorten diagnosis-to-action loops. Industrial deployments demonstrate up to 20-30% reduction in unplanned downtime when DT-enabled predictive maintenance is integrated into operational workflows Finally, we surface open challenges - data governance, model validation, uncertainty quantification, interoperability, and work-force adoption - and propose a practical roadmap to make DT predictive maintenance projects production-ready. 2025 IEEE. -
Emotion-Aware Sign Language Recognition Using CNN and UNET Architectures
This paper proposes an AI-based system for the recognition of sign language with the detection of emotions for more expressive communication among speech-impaired and hearing individuals and others. Traditional sign language systems focus mainly on the aspect of hand gestures and neglect the signs for emotions that add meaning and context. In order to overcome the limitation, the project proposes a system that utilizes Convolutional Neural Nets (CNNs) for the recognition of hand signs and UNET for the segmentation of the picture so that the area of the hands can be discriminated from the background. Facial Emotion Recognition (FER) is also incorporated in order to detect signs such as happiness, sadness, or anger. Overall, the parts together constitute a multimodal recognition system that can read signs and emotions and produce more natural and expressive outputs. The paper delves into architecture, dataset challenges, and implementation concepts with publicly available databases such as RWTH-PHOENIX-Weather 2014T. The combined approach can enhance inclusivity and access in learning, communication, and assistive technology. 2025 IEEE. -
Emotion-Aware Sign Language Recognition Using CNN and UNET Architectures
This paper proposes an AI-based system for the recognition of sign language with the detection of emotions for more expressive communication among speech-impaired and hearing individuals and others. Traditional sign language systems focus mainly on the aspect of hand gestures and neglect the signs for emotions that add meaning and context. In order to overcome the limitation, the project proposes a system that utilizes Convolutional Neural Nets (CNNs) for the recognition of hand signs and UNET for the segmentation of the picture so that the area of the hands can be discriminated from the background. Facial Emotion Recognition (FER) is also incorporated in order to detect signs such as happiness, sadness, or anger. Overall, the parts together constitute a multimodal recognition system that can read signs and emotions and produce more natural and expressive outputs. The paper delves into architecture, dataset challenges, and implementation concepts with publicly available databases such as RWTH-PHOENIX-Weather 2014T. The combined approach can enhance inclusivity and access in learning, communication, and assistive technology. 2025 IEEE. -
Railway Track Crack Detection: A Comparative Study On Yolov7 And U-Net In Automated Inspection
For railway networks to remain operationally safe and avoid catastrophic failures, structural integrity is essential. Track cracks can be found using labor-intensive, slow, and human error-prone manual inspection techniques. In this work, two cutting-edge deep learning models - YOLOv8 andU-Net v2 - for automated railway track crack detection using high-resolution imagery from Unmanned Aerial Vehicles (UAVs) are compared. In a real-world inspection scenario, we compare the different strategies of precise semantic segmentation (U-Net) and real-time object detection (YOLOv8) in order to assess their relative trade-offs. We compare performance on important metrics such as precision, recall, intersection over union (IoU), and inference speed using a custom dataset that was taken by a DJI Matrice 300 RTK drone. This work is novel because it examines how each model's output - bounding boxes versus pixel-level masks - directly affects the usefulness for maintenance workflows from an application-focused perspective. According to our research, U-Net v2 offers the fine-grained information required for precise damage assessment, while YOLOv8 is best suited for quick, extensive screening. This study offers railway operators useful information for creating a multi-stage, hybrid inspection strategy that strikes a balance between accuracy and speed. 2025 IEEE. -
A Context-Aware Finite State Machine for Gesture-Driven UAV Control
Gesture based Unmanned Aerial Vehicles (UAVs) is a very intuitive way to control drones (UAVs). Current methods tend to associate one gesture to one action, a practice which is rigid and inflexible. In this paper, we propose the Finite State Machine (FSM)-based gesture control framework, which allows triggering several actions of the UAV with a single gesture depending on the current state of the drone. MediaPipe hand gestures recognition and integration with ROS2 and PX4 allows the system to automatically takeoff, land, hover, automated ascents, and directional speed variation. Experiments in a ROS2 simulation environment test the system in terms of gesture-to-state latency, the rate of successful commands, the extent of the FSM that the system is capable of controlling, and the rate of false positives (spurious transitions). The results indicate that the suggested method has robust and responsive control of the UAV, which forms the basis to establish more intuitive and adaptive human-UAV interaction in limited spaces. 2025 IEEE. -
Enhancing Food E-Commerce Through Immersive Virtual Reality: An Reality: An Extended Technology Acceptance Model Approach for Consumer Adoption in the Post-Pandemic Era
Food purchasing differs from other types of internet shopping. With the introduction of the new retail structure, nearly every e-commerce platform has set up fresh food retail one after another. As a result, electronic gadgets have evolved into tools that marketers may use to initiate interactions with customers. Brands may use augmented reality enabled mobile applications to deliver precise information about products and services while also influencing consumer impressions. Perceived usefulness was the only factor that supported perceived ease of use as a mediator. Our findings provide useful information for researchers and industry experts to improve the effectiveness of VR systems by better understanding user adoption. 2025 IEEE. -
Adaptive Risk-Aware Ride Assignment (ARARA) Algorithm to Improve Efficiency to Lower Cancellation Rates in Bengaluru
Ride cancellations on urban mobility platforms like Rapido, OLA, Uber and other service provide platforms are negatively impacting user experience, driver earnings, and platform efficiency due to high cancellation of rides. This study addresses the challenge by developing a machine learning based adaptive userride matching algorithm that is trained on real world ride dataset from Bengaluru. The dataset includes features such as ride time, source, destination, distance, fare, payment method, and ride status. Through data preprocessing and feature engineering, key patterns influencing ride cancellations are identified. A classification model is developed to predict the likelihood of cancellation before ride assignment by using few Machine learning models among various model XGBoost and Logistic Regression outperformed with nearly 9 0% accuracy. Later to enhance the performance in allocation based on cancellation prediction the ARARA algorithm suggests that reallocates rides dynamically based on cancellation risk using inference and assignment Algorithm. Experimental results shows that how to reduce cancellation rates and improved accuracy by choosing best allocation based on top three best captains for allocation to optimize chances of cancellation. This framework can be integrated by ride platforms to enhance service reliability and optimize fleet efficiency. 2025 IEEE. -
Machine Learning-Based Credit Scoring for Personalized and Inclusive Lending in Consumer-Centric Financial Systems
Traditional approaches to credit-scoring are largely based on rule systems that can be excessively fixed and limited to the ability to reflect individual financial behavior. The article analyzes the effectiveness of machine-learning (ML)- based credit ranking with the hypothesis that they can improve predictive capability and fairness of consumer credit lending. The performance of these algorithms, including supervised methods of learning, e.g., logistic regression, random forests as well as the deep learning, is contrasted to the conventional credit models. Model transparency is provided by SHAP values and other methods explainable by AI. Findings show that practice based on the use of ML outperform traditional methods in risk assessment, especially, through the inclusion of supplementary forms of data in traditional databases based on transaction behavior, virtual footprints, and psychometric signals. Furthermore, ethical standards and moral confidence in ML informed credit decision-making will require regulation-proof and explanatory modelling. Through the research, it is recommended to implement policy measures intended to cause financial institutions, fintech companies, and regulating bodies to implement ML-based credit-scoring technologies, with fairness and predictive effectiveness being reciprocal drivers of financial access and consumer-friendly lending practices. 2025 IEEE. -
Cross-Platform vs. Native Mobile Development: A Comprehensive Study Using an Expense Tracker Application
In this paper, the Native Android, Flutter, and React Native are compared in an empirical way with the help of an Expense Tracker application on Android (Pixel 6, Android 14). The Native, Flutter, and React Native had an average start up time of 1.5 s, 2.4 s, and 3.0 s respectively. Maintainability Indexes and SUS scores demonstrate that Flutter provides efficiency without affecting the code quality. It was measured using standardized tools Android Studio Profiler, Trepn Profiler and SonarQube to measure performance, scalability and security. Findings show that Native Android provides the best performance and security, and Flutter is 35 times faster to develop and has a well-maintained code. Results guide the choice of frameworks by practitioners. 2025 IEEE. -
Smart UAV Surveillance Platform with Onboard Object Detection and Geofencing for Public Safety
Modern public safety operations are also changing due to the incorporation of artificial intelligence (AI) into the use of Unmanned Aerial Vehicles (UAVs) enabling maneuverability, open area, and continuous surveillance. The apparently exact plan and assessment of a clever UAV watchdog platform with on-board article locating, autonomous navigation using geofences, and edge processing will be introduced in this paper. The proposed system will be composed of a Pixhawk flight controller, a u-blox M10 GNSS module to provide navigation capabilities, and NVIDIA Jetson to run high-definition aerial video processing, specifically to recognize and use a YOLO-based detection pipeline. A simulated environment of a public event was used to test the accuracy of detection, inference latency, detection of anomalies and geofencing compliance across different environmental and crowd circumstances. The results of experiments also show average detection accuracy rates of 93.42%, inference delays of less than 60 ms, and few boundary violations, confirming the use of the system in responsive and secure use of aerial monitoring. The open-source modular nature of the platform permits the addition of other sensors and its subsequent expansion ability into multi-drone collaborative applications to note the prospects of such low-cost systems to influence large scale event tracking, disaster relief and urban security applications. 2025 IEEE. -
The Interaction of Generative Artificial Intelligence with Computational Intelligence on the Knowledge Economy: A Text Mining Approach
The study presents the first large-scale bibliometric evaluation on how the generative artificial intelligence (GenAI) and computational intelligence will impact the knowledge economy research during the year 2015 to 2025. Based on 228 articles indexed in Scopus, the study incorporated the use of the trend of keywords, multiple correspondence analysis and citation network visualization to determine the central thematic clusters and their trends in development. The findings revealed a notable growth of publications since 2020 due to the development of language-centered technologies such as large language models (LLMs), natural language processing (NLP) and generative adversarial networks (GANs), which have become the center of the intellectual cluster. Despite the diversity of research issues, only an estimated 9% of studies explicitly evaluate the role of GenAI in determining the consequences of the macroeconomic knowledge economy, which leaves a significant disparity between the pace of technological development and the overall effects of technology on the economy. This study provides a scholarly, practical oriented recommendation to use the power of GenAI to accelerate digital transformation and ensure equitable economic and social benefits. 2025 IEEE. -
A Compact Multi-Input DC-DC Converter for Smart Grid and Electric Mobility Applications
This paper presents the design and simulation of a compact multi-input DC-DC boost converter intended for renewable energy integration and electric mobility applications. The proposed architecture allows the simultaneous interfacing of photovoltaic panels, fuel cells, and battery banks to a common DC bus, enabling seamless energy management and hybrid source utilization. A maximum power point tracking (MPPT) algorithm is embedded to optimize PV energy extraction under varying environmental conditions. MATLAB/Simulink models are developed to analyze performance parameters including voltage gain, current ripple, power stability, and total harmonic distortion (THD). The converter achieves a stable DC output of 400 V from input ranges of 24-48 V while maintaining efficiency above 95%. Results confirm reduced current ripple (<50 mV) and THD below 3%, in compliance with IEEE standards for grid-connected power systems. The duty cycle modulation strategy minimizes diode voltage stress and optimizes transient response. Furthermore, grid synchronization tests validate the converter's capability for three-phase AC output with minimal distortion, making it suitable for smart grid integration and electric vehicle charging infrastructure. The proposed converter presents a viable alternative for sustainable and high-performance power electronic systems due to its compact size, reduced component count, and enhanced efficiency. 2025 IEEE. -
Comparison of Convergence Rates in Federated Learning and Federated Multi-Task Learning Using the CIFAR-10 Dataset
Many smart or cell phones have built-in distance, signal, and air pollution sensors. While collecting information, an acceleration registering device is a three-dimensional one and it can be applied in the gait analysis to address issues such as falls and health status determination. Indeed, the data is abundance in terms of quantity and some of the data may be of great concern in terms of privacy. In the time of Industry 4.0 the data has emerged as a key resource. Personal information/identity must not be maintained and hence cannot be stored at one place or all collected in a single place. AI models are moving to decentralized where a machine learning setting called Federated learning (FL) is being applied. FL has adversities such as statistical and systems heterogeneity. Actually, to better use shared information and build local models, Federated Multi-task learning (FMTL) has been devised. We also compare the number of iterations required to converge using CIFAR dataset of FL and FMTL. Several graphs illustrated in this paper show that convergence rates depend on the algorithm, number of communication rounds and number of clients or devices. Thus, it is clear that in some cases FL outperforms with FMTL in terms of convergence or conversely. However, it cannot be deduced that the type of FMTL always converges better than FL. The reliance on this graph is evident in this paper in order to as explain as prove the fact that, as the number of clients in FL rises, the rate of convergence declines. If ten communication rounds are employed with the use of the MOCHA algorithm, the model does not converge appropriately. The RMSE score declined from 1.14 to 1.02 throughout 20 epochs. 2025 IEEE. -
Mathematical Capabilities of LLMs
Large language models (LLMs) have the potential to solve mathematical problems well, but little has been investigated. In this research, we evaluate five leading LLMs - (Gemini, Claude, Mistral, ChatGPT, and Llama - on a set of 50 mathematical problems that cover calculus, algebra, geometry, number theory, and probability. Finally, the study evaluates the accuracy of their solutions and gives the ability to assess their intermediate steps correctly. I created a primary dataset for comparison of the LLM performances and ranked the models according to how well they were able to solve those problems. It illustrates the shortcomings of present LLMs at reasoning and solving for mathematics and suggests what needs to be rectified in the LLMs. Future research will further refine this dataset and monitor the progression of LLM capabilities in solving more complex mathematical problems. 2025 IEEE. -
Analyzing Financial Metrics: A Comparative Study of Salesforce and Microsoft Dynamics 365
The purpose of the study is, therefore, in the integration of six variables consisting of closing prices, daily returns, volatility, stock prices together with moving averages, trading volumes, and their moving averages: how these measure interrelates and what impacts they have toward market sentiment. Correlation and regression tests were carried out to ascertain whether the obtained findings were reliable enough to present robust relationships among the metrics of concern. Therefore, the obtained findings have wider implications for investors, as well as their investing decisions. In the first place, these findings permit one to detail and analyze in an in-depth manner the stock shares' behavior overtime and variations of the market so that trends within trading activity could be unearthed and understood long-term effects. 2025 IEEE.