Environmental, Social, and Governance (ESG) indicators have become critical determinants in sustainable investment and regulatory decision-making. However, building models that deliver both high predictive accuracy and interpretability remains a major challenge. This paper presents an explainable ensemble learning framework that enhances the interpretability of ESG score prediction through rule extraction and pruning. The proposed approach applies a structured preprocessing, feature encoding, and Random Forest-based ensemble learning. Further, decision rules are extracted and locally pruned to eliminate redundancy while preserving predictive power. The refined rules are aggregated using a weighted voting mechanism for multi-class ESG classification. Experimental evaluation on firm-level ESG datasets demonstrates that the proposed approach achieves competitive predictive accuracy (AUC $=0.90$, Precision $=0.72$) while producing concise, humaninterpretable rules. enhancing transparency and actionable insights for sustainable finance.
A distribution transformer (DT) is an intrinsic element in the overall power system. With increasing energy demands, the system requires an effective forecasting method at the distribution transformer level. With the recent emergence of deep learning models and their adaptation to this task, forecasting has generally improved; however, when incorporating weather and load features, these models often overlook granular electrical parameters that could enhance prediction reliability. In this paper, this gap has been addressed by incorporating overlooked electrical parameters alongside conventional features. Afterwords, state-of-the-art deep learning models have been deployed, including the Recurrent Neural Network (RNN), which is beneficial for sequence modelling; the Bidirectional LSTM (Bi-LSTM), which can capture dependencies in both forward and backward directions; and the hybrid RNN-Bi-LSTM, which increases computational efficiency by incorporating bi-directional learning. To further extract both spatial and temporal features, the CNN-Bi-LSTM model is utilized, which leverages its deeper architecture to extract features, thereby improving pattern recognition. An extensive evaluation of proposed models was conducted on a dataset of 48,000 instances by MIRAD LESCO using standard performance metrics, including MSE, RMSE, MAE, MAPE, and R2. The results have been compared with other notable studies in the domain and indicate the superiority of the proposed method, specifically RNN-Bi-LSTM, which achieves a MAPE of 9.476% and captures both short- and long-term dependencies, thereby enhancing predictive accuracy for ShortTerm Load Forecasting (STLF) of distribution transformers. The proposed work enhances forecast accuracy at the DT level, enhancing load management and operational efficiency.
Digital twins integrate analytics and human expertise to create models that support informed healthcare decisions. In resource-limited areas, access to medical care is often limited by a lack of facilities, skilled professionals, and transportation. Wireless capsule endoscopy (WCE), being portable, reliable, and easy to use, has emerged as a superior alternative to traditional endoscopy, particularly for patients in remote areas. However, WCE generates large volumes of video data, requiring significant computation to analyze and extract relevant information. To address this challenge, we propose a video summarization scheme that identifies key frames while removing redundant content, preserving essential diagnostic information. Furthermore, we introduce a deep learning-based digital gastroenterologist twin for the automated classification of stomach-related pathological findings. The system reduces storage requirements without compromising critical information, enabling gastroenterologists to provide remote support. Experimental results demonstrate a performance improvement of at least 3% compared to state-of-the-art deep learning techniques, highlighting its potential for intelligent, scalable, and efficient gastroenterology care.
Approximately 1.13 billion websites span diverse domains like music, movies, health, video games, social media, entertainment, and more, making website classification highly valuable yet challenging. This research explores using Artificial Intelligence techniques to categorize websites into appropriate classes. The DMOZ dataset is one of the most reliable data sources for research focused on URLs, containing about 1.5 million URLs manually classified into different categories, though the distribution of URLs across different categories is highly unbalanced. To overcome this limitation, we collected 500,000 URLs from DomCop's Top 10 Million Websites and classified them using Large Language Models (LLMs), specifically LLaMA 3 3B. Numerous machine learning and deep learning models were trained on both datasets, including logistic regression, support vector machine, decision trees, random forest, CatBoost, XgBoost, LSTM, Bi-LSTM, and FastText. Different data handling and pre-processing techniques were also employed, including chunking for balancing and subdomain extraction to increase URLs per category. We report the most successful pipeline using the FastText model, achieving an average accuracy of 82%.
Software bugs have remained a prominent problem in contemporary software development environments, resulting in financial losses, project delays, and system unreliability. Conventional bug prediction models have proven less effective when dealing with higher-dimensional problem spaces and less relevant attribute information. To counter these effects, this paper proposes a hybrid machine learning model incorporating PCA, Mutual Information, RFE, and an Ensemble Boosting classifier. In this hybrid model, PCA is used to capture the principal directions of variance, which can successfully handle dimensional issues. Mutual Information and RFE are used in combination to discover and select more informative and discriminating features. Ensemble Boosting is used as a classifier to successfully learn non-linear relationships among these informative features. The hybrid model proposed in this paper performed defect accuracy tests on a variety of NASA datasets, with an accuracy of 97.
The growing complexity of network environments and the rapid emergence of zero-day attacks have undermined the effectiveness of traditional signature-based intrusion detection systems (IDS). To address these limitations, this paper proposes an autoencoder framework, termed a Residual Autoencoder (RAE), for unsupervised anomaly-based intrusion detection. The proposed model is trained solely on benign traffic and detects abnormal behavior by computing reconstruction errors on unseen data. This is achieved by adding a residual skip connection to maintain information of low-level features and stabilize gradient flow during training. Furthermore, a hybrid loss, which is a combination of Smooth L1 loss and cosine similarity, simultaneously applies numerical precision and structural consistency between the input and the reconstructed features. The experiments using the UNSW-NB15 dataset show that the proposed RAE is an effective framework to differentiate between normal and attack traffic, with an ROC-AUC score of 0.903, surpassing baseline methods, including the plain autoencoder (0.884), isolation forest (0.860), and one-class SVM networks (0.820). The results indicate that the model can effectively handle new and zero-day attacks. Additionally, it offers a robust, data-efficient, and scalable solution for next-generation network intrusion detection systems.
Video surveillance has become an essential part of modern safety systems. However, most of the existing solutions depend on large models that require significant computational resources and raise privacy concerns, making them less suitable for domestic environments. To address these limitations, this work presents SmartGuard, an edge based intelligent surveillance framework designed to detect suspicious activities in real time using low cost hardware. Its ability to upgrade conventional CCTV systems into intelligent surveillance systems makes it highly practical for home use. The system operates entirely on an NVIDIA Jetson Nano and employs a two stage pipeline: YOLOv8n detects relevant objects, while a fine-tuned MobileNetV2 classifier identifies activity types. Transfer learning was applied to a modified UCF-Crime dataset, combined with a newly collected fence climbing category comprising 11,000 frames. SmartGuard triggers alerts only when at least 70% of frames within a 3second window are classified as suspicious, effectively minimizing false positives. Experimental results show that the fine-tuned MobileNetV2 achieves 73% accuracy and runs reliably at 8-15 FPS on the Jetson Nano. These findings demonstrate that SmartGuard offers a privacy preserving and cost effective edge solution for real time domestic surveillance. The project codebase-related content is available at https://github.com/INFILP/SmartGuard-CodeBase-And-Results
The exponential growth of online banking has dramatically increased the potential for and complexity of fraudulent operations. Particularly with skewed and high-dimensional datasets, traditional fraud detection techniques often fail to adapt to developing trends. Our research presents a new financial transaction anomaly detection approach using a hybrid deep learning framework that combines Autoencoders, Convolutional Neural Networks (CNNs), and Long Short-Term Memory (LSTM) networks. The model employs CNNs to extract spatial features, LSTMs to capture temporal relationships, and autoencoder-based reconstruction errors to detect unobserved fraudulent behavior. Precision, recall, F1-score, and AUC-ROC were measured on a real-world credit card fraud dataset, and the results show that the hybrid model performs far better than the conventional machine learning and solo deep learning baselines. We use Shapley Additive Explanations (SHAP) to ensure the model can be understood, allowing for precise examination of the choices made after the fact. In addition to functioning well in highly unequal class situations, the findings show that the framework satisfies the essential transparency standards for use in regulated financial settings.
Conventional, isolated security models fail to protect IoT-integrated Smart Grids from combined cyber-physical threats. Using a co-simulation testbed, this research proves that cyberattacks directly cause physical damage. In this paper, the need for holistic analysis with the help of a high-fidelity cosimulation testbed (NS-3, GridLAB-D, HELICS) is proved. We demonstrate that cyberattacks have a direct impact on causing critical physical failure; an example is a false data injection attack that leads to the severity of voltage and line loading violations. More importantly, we demonstrate a synergistic compound effect: a Denial-of-Service attack, which paralyzes situational awareness of more than 45 percent packet loss, proactively blocks the detection of the FDI-induced violations. This covering effect dramatically enhances the threat of domino failures- an amplifier of risks to which the current siloed models are incapable of capturing. Our effort offers an established avenue on which effective countermeasures to these joint threats can be built.
Access to quality healthcare remains a persistent challenge in remote and underserved regions due to infrastructural limitations, intermittent internet connectivity, and a shortage of trained medical professionals. This study proposes and evaluates a lightweight edge computing framework integrated with Artificial Intelligence (AI) to address these barriers and enable real-time, offline diagnostics at the point of care. The system deploys optimized AI models on low-power devices, allowing local inference of patient vitals and symptoms without relying on cloud connectivity. Tested in a simulated rural environment using Raspberry Pi and Edge TPU platforms, the framework demonstrated significantly lower latency, reduced data transmission, and improved system reliability compared to conventional cloud-based solutions. Usability testing confirmed the framework's accessibility and effectiveness, even for non-technical users. The study highlights the potential of edge-AI systems in enhancing healthcare access, ensuring data privacy, and supporting scalable health delivery models in developing countries. Future directions include the integration of federated learning, expansion into maternal and mental health diagnostics, and mobile edge deployment for crisis response and policy integration.
The rapid growth of Intelligent Transportation System (ITS) leads to continuous generation of vehicular data that must be processed under latency and connectivity constraints on heterogeneous computing infrastructures (e.g., Vehicular Edge Computing (VEC), Mobile Edge Computing (MEC), and Cloud Computing (CC)). Under these conditions, task offloading must account for mobility dynamics and server status to efficiently manage communication delay, cost, and computational resources. To address this challenge, this work introduces a mobilityaware offloading mechanism using Q-learning, which decides on the optimal tier. The decision process incorporates real-time indicators, including server utilization, signal quality, and Base Station (BS) handover patterns, allowing the scheduler to adapt to changing environments proactively. The proposed solution favors VEC when sufficient resources exist, shifts to MEC during VEC over-utilization, and selects CC only as a last option. A customized reward model guides the learning agent toward resource-efficient policies while penalizing unnecessary tier switching. The approach is implemented and assessed using DriveNetSim, a specially built vehicular simulator that realistically emulates mobility, signal attenuation, and BS switching. Evaluation results show that the proposed method achieves up to 43% reduction in communication delay and 38% lower processing cost, demonstrating its suitability for highly dynamic vehicular networks.
Thyroid cancer is an emerging global health concern that requires early and accurate diagnosis to improve patient outcomes. Conventional diagnostic techniques, such as ultrasound imaging and fine-needle aspiration biopsy (FNAB), suffer from inter-observer variability and subjectivity. Deep learning (DL) has demonstrated strong potential for automated thyroid cancer diagnosis with high accuracy and efficiency. This paper presents a comparative study of state-of-the-art convolutional neural networks (CNNs), including ResNet, DenseNet, and EfficientNet, and transformer-based architectures, including Vision Transformer (ViT), ConvNeXt, and Swin Transformer, for thyroid nodule classification. Model performance is evaluated using accuracy, precision, recall, and F1-score, along with computational efficiency for clinical deployment. To enhance transparency and clinical trust, explainable AI techniques such as LIME and Grad-CAM are applied to visualize salient regions in ultrasound images. Experimental results indicate that transformer-based models achieve the highest accuracy of 96.3%, while EfficientNet offers a favorable trade-off between performance and model complexity. Finally, challenges and future research directions for integrating AI-based diagnostic systems into clinical practice are discussed.
Glaucoma is a neurological condition that causes progressive damage to the optic nerve. The major goal of this study is to develop automatic ocular image analysis based on deep learning to aid medical diagnosis in glaucoma-related eye disorders. First, VGG16 model is implemented for glaucoma detection and classification. The second model that has been employed in this research is Residual network (ResNet50). The third model is Inception-V3, and the final model is an ensemble model. An ensemble model combines the outputs of implemented classifiers to create a composite classification. The idea of combining all the classifiers is to create a hybrid model that outperforms each individual classifier in terms of classification accuracy. It identifies the four stages of glaucoma disease: advanced, early, moderate, and normal. The ensemble model is proposed and implemented a Deep Convolutional Neural Network for glaucoma detection and classification. The model was fine tuned to improve the classifier. This proposed system had an excellent ability to classify normal and different stages of glaucoma with an accuracy of 98.42%. We found ensemble model is a suitable deep learning method to classify normal and different stages of glaucoma.
Wireless power transfer (WPT) has emerged as a transformative technology in electric vehicle (EV) infrastructure, enabling contactless charging that enhances convenience, safety, and system resilience. However, efficiently managing energy flow within hybrid storage systems, comprising electrochemical batteries and ultracapacitors, remains a critical challenge due to their distinct electrical dynamics. Batteries provide high energy density for sustained operation but are vulnerable to rapid current fluctuations, while ultracapacitors offer exceptional power response with limited energy storage capacity. This paper describes an AI-enabled power distribution system which combines rule-based control systems with a generative pre-trained transformer (GPT3.5) model for the real-time and situational energy management of hybrid storage systems with Wireless Power Transfer (WPT). The controller intelligently steers transient power where it needs to be to an ultra capacitor while maintaining battery health by controlling the battery's charge/discharge cycles. The improvements of the AI system on the control system's energy transfer efficiency and diminished rate of electrochemical aging is verified in the simulations and real-time computational tests. The system's adaptability to variable levels of wireless power and load is also noted. The projected outline achieves efficiency optimization by 8% and 50% less battery stress compared to rule-Based control while laying the groundwork for automatic environmentally friendly wireless EV charging infrastructure.
Air pollution is an ever-increasing threat to human beings and the environment, particularly in big urban centers. Numerous of the existing Air Quality Index (AQI) monitors fail to make predictions, sensor calibration is not reliable, and they are not updated in real time. This paper presents an IoT-based AQI monitoring and forecasting system, which is based on a Transformer deep learning model to implement timebased predictions and cheap, well calibrated sensors. This system captures weather and environmental data in real-time, performs feature engineering and normalization, using timestamps to compute and predict the AQI value of up to 48 hours. Good predictive performance is demonstrated by experimental results, and the R2 score is 91.99. The framework assists smart cities in making data-driven decisions by providing a forecasting system of air quality which is accurate and scalable.
As social media has grown, some people with biased opinions now use it to spread hateful messages and target others because of their gender, race, or ethnicity. Positive communication in diverse communities can build confidence, but negative remarks can damage a person's image and emotional health. The increasing use of social media shows the need for reliable systems to detect and reduce harmful content. In this study, we utilized classical machine learning methods for cyberbullying detection, applying three word vectorization techniques; TFIDF, Word2Vec (Skip-Gram and CBOW), and FastText. These were combined with five traditional machine learning classifiers. Among all models, Logistic Regression and Support Vector Machine integrated with supervised FastText embeddings achieved the highest accuracy of 0.97. Finally, we used Captum to explain and visualize how the model makes its decisions. This helps people better understand the results and makes the model more transparent.
E-governance enables improved service delivery to citizens through communication technology that is more efficient, interactive and transparent as compared to paper-based record keeping. Pakistan's judiciary has initiated a process of digitalization over the last decade and a number of online portals have been established to allow people to search for the status of their cases. The availability of these case management systems offer an opportunity to improve transparency, accessibility, and data-driven decision making. However, the quality of the data available on these online portals has not yet been investigated. For this study, we have extracted and analyzed case-level data from two high courts of Pakistan: the Sindh High Court and the Punjab High Court. We used web-scraping techniques to compile the data, and more than 11,000 entries for the Sindh High Court, and around 1,000 entries for the Lahore High Court were extracted. The data is evaluated for accessibility, consistency, completeness, and usability. Our findings suggest that data has inconsistency issues including missing values, redundant information, and limited historical coverage. We found that despite ongoing digitalization, both portals lack standardization, hindering comprehensive legal research. Such data related issues may be one of the reasons for the limited quantitative research on our judicial system. This study is an attempt to provide a structured dataset to establish a baseline assessment that can support future computational research and evidence-based policy making.
Radio Frequency Identification (RFID) systems are widely deployed in access control, inventory tracking, and track-and-trace applications due to their low cost and ease of integration. Ultra-lightweight authentication protocols are typically used to ensure the security and privacy of the communication channel between the tag and reader pair. This paper presents the security analysis of a recent protocol, i.e, Ultra-lightweight RFID Authentication Protocol (URAP). The proposed work highlights a critical vulnerability in the architecture and working of the URAP protocol. To mitigate the threat, we propose a lightweight enhancement that introduces timestamp validation into the authentication process, improving resilience against denial-ofservice scenarios without compromising protocol efficiency.
A unified approach for the Nonlinear Model Predictive Control (NMPC) is presented to control active front steering, lateral acceleration, and longitudinal velocity for a rear-wheel-driven autonomous electric vehicle. The purpose is to ensure reference tracking and system stability by utilizing multi-objective optimization functions. Through the results, the effectiveness of the proposed approach is demonstrated to achieve the robustness and stability. The dynamic model of a rear-wheeldriven test vehicle also incorporates the Pacekja tire model, that is considered as the most accurate tire model. The performance of the proposed controller is evaluated for the constant-radius maneuver test under a dry asphalt road surface for the rear-wheel-driven EV. Through simulation results, it is observed that the NMPC is able to successfully maintain longitudinal velocity, lateral acceleration, while following the desired trajectory.
