Heart disease remains a leading cause of death globally, underscoring the need for early detection to improve outcomes and reduce healthcare cost. This study proposes a machine learning-based approach to predict heart disease risk using a comprehensive dataset covering demographic, lifestyle, and clinical variables. The study addresses limitations in prior research by focusing on fairness, interpretability, and practical deployment for clinical utility. The methodological framework involved stratified data splits and careful oversampling to prevent leakage. The anonymised Kaggle/BRFSS dataset was partitioned into 70 % training, 15 % validation and 15 % test sets, and SMOTE oversampling was applied only on the training split to balance the minority (heart disease) class. Feature selection via mutual information highlighted HadAngina, BMI and AgeCategory. Hyper-parameter-tuned models-including Logistic Regression, Random Forest, XGBoost, K-nearest neighbour and multilayer perceptron-were trained using stratified 5-fold crossvalidation with fixed random seeds; XGBoost achieved the highest performance (accuracy 91.99%, ROC-AUC 0.9756). Fairness was assessed using demographic-parity difference, equal-opportunity (true-positive-rate) gap and calibration error across sex, age and the official race/ethnicity categories (White, Black, Asian, Hispanic and Other). Initial results revealed lower recall for younger and Black individuals. A simple re-weighting of the training data and group-specific threshold adjustments reduced the demographic-parity difference to 0.0875 with minimal impact on accuracy. SHAP values provided global and local explanations of feature contributions-crucial for building trust among clinicians and patients. The model was deployed via a Flask-based web app, providing real-time heart disease risk assessments with user-friendly interfaces and explanations for clinicians and the public. This study bridges AI and healthcare by delivering a fair, transparent, and accessible predictive tool. Future enhancements include adversarial debiasing, wearable data integration, and expanding to other cardiovascular conditions to improve diagnostic accuracy, reduce inequities, and increase trust in AIdriven healthcare.

This study examines how firm size shapes performance and risk-taking among FTSE 100 firms from 2018-2023, a period spanning Brexit and COVID-19 market turbulence. Using daily market data, performance is defined by next-day stock price direction, firm size by the logarithm of market capitalization, and risk-taking by rolling return volatility. A comparative supervised machine-learning framework evaluates Logistic Regression, Support Vector Machines, Random Forest, Extra Trees, Decision Trees, Gradient Boosting, and K-Nearest Neighbors on scaled features that include volatility, leverage, and capital intensity, with cross-validated out-of-sample tests. The main contribution is UK-specific evidence that a regularized linear classifier (Logistic Regression) achieves state-of-the-art accuracy while preserving interpretability for corporate finance decisions, outperforming more flexible non-linear models on noisy, high-frequency equity data. Logistic Regression attains the top performance (97 % accuracy; ROC-AUC 0.98), with SVM close behind; tree ensembles consistently rank volatility as the most influential predictor of performance. Empirically, larger firms exhibit more stable returns and lower unstructured risk, consistent with superior risk management capacity. Practically, the results provide a reproducible screening pipeline for FTSE 100 equities by risk profile and size, and theoretically they indicate that the dominant decision boundary in this feature space is near-linear.

The integration of modern authentication methods such as multifactor authentication (MFA), biometric verification, and password-less authentication in legacy systems presents significant challenges due to outdated infrastructure, compatibility issues, and security vulnerabilities. This study analyzes the difficulties associated with implementing these authentication techniques in legacy environments and proposes viable solutions. A combination of software-based adaptation layers, API gateways, and Zero Trust frameworks can help bridge the gap between traditional and modern security paradigms. The research also includes algorithmic approaches for retrofitting authentication mechanisms, security models for risk assessment, and implementation case studies. The results indicate that a hybrid approach combining incremental upgrades with identity federation mechanisms enhances authentication security while minimizing operational disruptions. The study concludes that legacy systems require a strategic balance between modernization efforts and system stability, ensuring long-term security and efficiency in authentication processes.

This study assesses four machine learning (ML) algorithms – Decision Tree (DT), Random Forest (RF), K-Nearest Neighbours (KNN) and Support Vector Machines (SVM) – for predicting customer churn in banking. It explores optimisation methods including SMOTE, hyperparameter tuning and feature importance to enhance model effectiveness. Results indicate that RF outperforms the others, showing significant metric improvements over previous studies.

This study investigates task allocation optimization in IT project management using cooperative game theory, focusing on the Shapley Value approach. Effective task allocation is critical in IT projects to balance technical proficiency and team cooperation. By modeling developers and tasks within a cooperative framework, this research contributes to the field by proposing a systematic method to quantify individual contributions and maximize team performance. Through a Python-based implementation, the study demonstrates the scalability of the approach and evaluates its effectiveness in various scenarios, focusing on fairness and efficiency.

Chest radiographs (CXRs) remain one of the most widely used imaging modalities for diagnosing cardiopulmonary diseases, yet increasing imaging demand and a shortage of radiologists continue to challenge diagnostic accuracy and efficiency. Deep learning approaches offer promising solutions, but their clinical adoption is hindered by limited interpretability. In this study, we compare a baseline convolutional neural network (CNN) augmented with Gradient-weighted Class Activation Mapping (Grad-CAM) to a multimodal framework that integrates large language models (LLMs) for explainable CXR interpretation. The baseline CNN, trained on the ChestX-ray8 dataset of 108,848 images across 14 thoracic pathologies, achieved a mean Area under Curve (AUC) of 0.891 and macro F1-score of 0.80, with Grad-CAM heatmaps providing visual explanations of model predictions. In contrast, the multimodal framework combined CNN-derived features with clinical narratives using cross-attention and fewshot prompting and was fine-tuned on 68,000 paired image-report samples. This approach improved diagnostic performance (AUC=0.915, F 1=0.89), generated human-readable reports, and localized anatomical features to enhance interpretability and clinical reasoning. Our findings demonstrate that multimodal AI substantially improves precision and transparency over CNN-only systems, offering a more reliable decision-support tool to assist radiologists, reduce diagnostic errors, and optimize workflow in high-volume clinical environments.

This project develops a blockchain-based web application aimed at verifying pharmaceutical products, a critical step in combating counterfeit drugs. By leveraging Ethereum's Sepolia test network and the power of smart contracts, the system facilitates secure, transparent processes for registering, verifying, and tracking pharmaceutical items. The application combines a PHP-based backend with a JavaScript-powered frontend, seamlessly integrating tools like MetaMask for user authentication and Web3.js to enable blockchain communication. The study underscores the significant advantages blockchain offers over traditional verification methods, particularly in terms of data integrity, transparency, and security. As a result, it meets the CIA triad model's requirements for confidentiality and integrity in information security. The novel features of the project include the use of optimism roll-ups for scalability, two-factor authentication (2FA), and data encryption to address critical challenges often overlooked in similar proposals for blockchainbased authentication systems. The results of the project demonstrate a tangible improvement in supply chain transparency and fraud prevention within the pharmaceutical industry. This work lays a strong foundation for further exploration of decentralised applications, not only in pharmaceutical validation but also across other essential sectors.

This study presents the design and development of a real-time, machine learning-based human activity recognition (HAR) system. The research focused on building a Convolutional Neural Network (CNN) model integrated into a web-based application for real-time detection and classification of human activities. Four experimental setups were implemented: a baseline CNN, CNN with data augmentation, VGG16 with transfer learning and augmentation, and Inception ResNetV2 with augmentation. The models were trained and validated using a labeled human activity image dataset from Kaggle. Model performance was evaluated using accuracy, precision, recall, and computational cost. The models were implemented using TensorFlow and Keras in Python, and deployed via a Flask-based web application for real-time activity detection. Among all models, the Inception ResNetV2 with data augmentation achieved the best performance (accuracy$=0.762$, precision$=0.84$, recall$=0.69$), demonstrating strong generalization and suitability for real-time applications. The novelty of this work lies in integrating an optimized deep learning model with a web-based real-time processing framework, offering an efficient and scalable solution for activity recognition in domains such as healthcare and sports.

This paper presents a real-time assistive system designed to support visually impaired individuals through object detection, voice-guided navigation, and speech-based information retrieval. The system integrates the YOLOv5s deep learning model for object recognition and a speech interface for intuitive auditory feedback. The proposed model was implemented on a system powered by an Intel(R) Core (TM) i5-8250U CPU @ 1.60$\mathbf{G H z}$(up to 1.80 GHz) with 8 GB RAM and integrated Intel(R) UHD Graphics$\mathbf{6 2 0} \boldsymbol{(} \mathbf{1 2 8 ~ M B} \boldsymbol{)}$, utilizing the built-in system camera for real-time video input. Quantitative evaluation demonstrated an object detection accuracy of$89.6 \% \text{mAP}$, a navigation success rate of 96%, and a speech recognition accuracy of 93%. The system achieved an average latency of 0.72 s for object detection, 1.1s for navigation, and 0.68 s for voice retrieval, with a frame rate of 22-25 FPS on GPU hardware. Performance testing was conducted under varying lighting conditions, partial occlusions, and different noise levels$(40-70 ~\text{dB})$to evaluate environmental robustness. The results confirm that the system maintains stable operation in both indoor and outdoor environments, demonstrating practical applicability and real-time responsiveness for assistive navigation in dynamic and crowded settings.

Financial fraud, particularly in credit card transactions, remains a critical issue for financial institutions due to the growing complexity of digital payment systems and cyberattacks. This study investigates the effectiveness of five machine learning algorithms-Logistic Regression, Naive Bayes, Random Forest, Light Gradient Boosting (LightGBM), and Extreme Gradient Boosting (XGBoost) and these were selected for their complementary strengths in interpretability, scalability, and ability to handle nonlinear relationships. A deep learning Autoencoder was also used as an anomaly detection baseline for fairness of comparison. The dataset was highly imbalanced (1% fraud), addressed using SMOTE resampling and MinMaxScaler normalization. Expanded feature analysis, including descriptive statistics, correlation patterns, and feature importance, revealed that transaction amount and merchant category were the strongest predictors of fraud. Experimental results show that Random Forest outperformed all other models, achieving 99% accuracy, 0.99 recall, and a perfect ROC-AUC of$\mathbf{1. 0 0}$. These findings demonstrate that ensemble-based methods are both accurate and stable for real-world fraud detection. Practical implications regarding model interpretability, deployment, and fairness are also discussed. The findings of this research have practical implications for financial institutions looking to implement automated fraud detection systems, ultimately improving their ability to minimize financial losses and enhance the security of digital payment systems.

Brain tumors, a significant health concern due to their potential to disrupt critical brain functions, require accurate and timely detection for effective management. Traditional diagnostic methods, reliant on manual examination of MRI scans by radiologists, often face challenges related to time constraints and susceptibility to human error. This study investigates the application of deep learning techniques to automate the detection and classification of brain tumors from MRI scans. We evaluated various neural network architectures, including CNN, VGG16, VGG19, and ResNet-50, to determine their effectiveness in identifying and classifying brain tumors from MRI images. Our results demonstrate that the CNN model outperforms VGG16, VGG19, and ResNet-50 in terms of both accuracy and generalization, making it the most effective choice for automated tumor detection. To address practical clinical needs, we developed a user-friendly web application that integrates the CNN model, enabling real-time tumor detection and classification. This application allows healthcare professionals to upload MRI images and receive immediate tumor detection and classification results, facilitating quicker and more precise diagnostic processes. The integration of deep learning models into this web-based platform marks notable progress in the automated detection of brain tumors. By enabling realtime analysis, the application supports clinicians in making informed decisions and planning treatment strategies with enhanced accuracy.

The global film industry faces significant challenges in making content accessible across diverse linguistic and cultural contexts, particularly through dubbing, which involves translating and synchronizing audio for different languages. Traditional dubbing methods, while effective, are labour-intensive, costly, and often struggle with preserving cultural nuances and emotional depth. With advancements in artificial intelligence (AI), there is a promising opportunity to address these challenges and revolutionize the dubbing process. This research explores the application of AI-driven technologies to improve the efficiency and accuracy of dubbing English-language films into Urdu. The research investigates the current state of AI-driven dubbing technologies, identifies their limitations, and proposes methods to enhance their effectiveness. Key objectives include analysing AI technologies for translation and speech synthesis, developing a streamlined process for English-to-Urdu dubbing, and evaluating the cultural and emotional fidelity of AI-generated content. A comprehensive methodology involving data extraction, speech-to-text conversion, translation, text-to-speech synthesis, and audio-video integration is employed to create an automated dubbing system. The system’s performance is assessed based on accuracy, naturalness, and synchronization, revealing both successes and areas for improvement. The findings demonstrate that AI-driven dubbing can significantly enhance the efficiency and scalability of the dubbing process, although challenges related to accent variability, cultural sensitivity, and voice naturalness remain. The study concludes with recommendations for future research, including the need for improved transcription accuracy, advanced translation models, and real-time dubbing capabilities. This research contributes valuable insights into the potential of AI to transform film localization and other sectors by making content more accessible and culturally relevant to diverse audiences.

This paper introduces Agri Sage, a mobile application aimed at addressing critical challenges in modern agriculture. Agri Sage integrates real-time plant disease detection using machine learning, an e-commerce platform for agricultural products, and weather updates alongside government pricing and subsidy information. Leveraging TensorFlow Lite for image analysis, the app diagnoses plant diseases even under suboptimal image conditions. The e-commerce platform allows farmers to connect with buyers and manage bulk transactions, while real-time weather and government data help farmers make informed decisions. This holistic approach enhances farm management, boosts productivity, and makes technology accessible to farmers in diverse environments. Extensive testing of Agri Sage demonstrated its efficacy in real-world applications, particularly in improving crop health and market access.

With the continuous advancement of technology and the need to keep pace with the digital era, the implementation of robust automated job recommendation systems has become essential to address the limitations of traditional methods and manual processes. This dissertation focuses on developing an online job search website application that integrates an automatic recommendation, alert and notification system, facilitating a more efficient connection between employers and job applicants using ML. Employers will have the ability to post job openings, review applicant profiles, and select the most qualified candidates. The user profile will be utilized to recommend jobs to candidates through a Semantic-Based Search System, with the Cosine Similarity technique serving as the key factor for automated job recommendations and comparing alongside the behavior of Jaccard similarity and the Jaccard similarity with subset matching. In other to understand what the best scenario for use for each is. The development of this job portal application aims to address the challenges faced by companies in filling vacancies and by job seekers in finding suitable employment opportunities.

The development of drowsiness detection applications has become increasingly important within the automotive technology sector, success rates are far from optimal. Despite the potential life-saving impact of such applications, the possible role of dynamic risk management in the successful completion of drowsiness detection projects has not been directly explored in existing studies on the topic. Considering the theoretical, as well as practical significance of the issue, the research under consideration aimed to analyze the benefits, challenges, and best practices of dynamic risk management in projects like “Drive Guard” a drowsiness detection Android application, with emphasis on the development of a framework for its implementation in such applications.

This research aims to develop machine learning (ML) models for the early identification of diabetes, a chronic condition that presents considerable global health threats, including heart disease, kidney failure, and neuropathy. Conventional diagnostic techniques, such as fasting plasma glucose (FPG) and HbA1c tests, tend to be invasive and often recognize the illness only in its advanced stages. To tackle this issue, the study investigates a range of ML algorithms, including Logistic Regression, Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Decision Trees, AdaBoost, Naive Bayes, and XGBoost. It also incorporates comprehensive data preprocessing methods like feature scaling and Synthetic Minority Over-sampling (SMOTE). The goal of these initiatives is to improve the models’ performance and interpretability to facilitate earlier diagnosis. Among the models assessed, K-Nearest Neighbors (KNN) and AdaBoost stood out, both achieving an accuracy of 0.7727, with XGBoost following closely at $\mathbf{0 . 7 6 6 2}$. After tuning, Support Vector Machines (SVM) showed a significant improvement, reaching an accuracy of 0.85, while KNN achieved an accuracy of 0.87 on the test set. Although Decision Trees and Naive Bayes performed less effectively, this research highlights the necessity of model interpretability in clinical settings. Techniques such as SHapley Additive exPlanations (SHAP) are utilized to elucidate predictions. The implementation of these models on a web-based platform using flask framework which allows for real-time assessments of diabetes risk and insights into critical risk factors, aiding healthcare providers and patients in making better-informed decisions.

With thousands of passengers travelling through airports on a daily basis during air travel, it is particularly important to ensure that the lives of passengers and crew are safe. In modern airport security screening, it is vital to ensure the efficient and accurate detection of threats in passenger baggage, which will have a bearing on whether or not the safety of company and individual property can be effectively safeguarded. In addition, failure to effectively improve the accuracy and efficiency of threat item detection during airport security screening may waste security resources and increase passenger inconvenience, leading to anxiety and dissatisfaction during screening and affecting their travelling experience. Furthermore, if the detection technology is not effective in screening threatening items, airports may need to increase the number of security personnel to manually check baggage, which can significantly increase operational costs. This research therefore utilises YOLOv5 object detection technology to improve the efficiency of airport security systems. YOLOv5's real-time capabilities can quickly identify potential threats by detecting prohibited items in baggage X-ray images. However, the black-box nature of deep learning models such as YOLO often prevents the transparency required for critical security applications. To address this issue, LIME (Local Interpretable Model- agnostic Explanations) was integrated into the workflow to provide interpretable visualisations that shed light on the decision-making process of YOLO models. By applying LIME, we generated visual interpretations that highlighted key areas of the image that influence model predictions, leading to a deeper understanding of model behaviour and increased trust in the system. The experimental results show that the accuracy of detecting Seal increased to 0.64, indicating a significant improvement in the model's ability to detect this category. The accuracy rate reaches ${1 . 0 0}$ at a confidence level of 0.936, implying that almost all predictions are correct when the model's prediction confidence level exceeds 0.936.The YOLO-LIME combined framework not only maintains high detection accuracy, but also significantly improves the interpretability of the model's decisions. This dual approach allows security personnel to better understand and validate model outputs, ultimately contributing to the reliability and transparency of airport security processes.

The rapid spread of fake content across digital platforms, including text, and images, poses significant challenges to the integrity of information. While recent advancements in multimodal fake detection have shown promise, existing models often focus on a single modality or lack scalability and transferability to unseen events. This study addresses these limitations by developing a novel fake detection model based on the CLIP and BLIP architectures, designed to be scalable across text and image modalities. The proposed architecture makes use of a BLIP model to extract image-text similarity and a CLIP model to extract image and text embeddings. The hybrid fusion technique was used to fuse the features. The proposed model was evaluated against existing bimodal approaches, demonstrating superior performance with an accuracy of 90.63%, precision of 92.04%, recall of 90.73%, and an F1-score of 91.38%. These results highlight the model’s effectiveness in accurately detecting fake content across diverse modalities while maintaining a balanced performance between precision and recall. This research contributes to the advancement of multimodal fake detection by providing a scalable and comprehensive approach, paving the way for future developments in combating misinformation across various digital mediums.

Daily retail sales are impacted by a lot of external factors, holidays and special events are one such category. In general, retail sales are largely impacted by fluctuations in demand hence, it is common for a retailer to run out of stock for some items and overstock the other items and this happens due to the lack of understanding of the actual number of items in demand for a particular item at a particular time of the month. This research work examines the impact of holidays or special events on the sales of a wide category of items using predictive analytics. It is done by performing exploratory data analysis and pre-processing methods followed by feature engineering and information extraction to extract the optimal input parameters to be fed into the model. The dataset is time-series data however, advanced machine learning algorithms are also used along with time-series methods to see if time-series data works well with non-time-series algorithms. Different time-series methods along with gradient boosting and the Facebook prophet model are evaluated in this work, achieving 92.83 % forecast accuracy with the Facebook prophet model. The gradient boosting model performs well with a MAPE value of 22.25% and time-series Holt Winters' additive method provides a MAPE value of 12.84 %. Each of the algorithms provides a good score with this time-series data and an appropriate algorithm can be chosen as per the business need.

Phishing website detection is the task of classifying websites as phishing or legitimate based on URL parameters and certain behaviour of the site. In today's world, dependency on websites has become inevitable. With the increase in website users population and the rise of the internet, cyber-attacks have become a common thing. Attackers across the globe target innocent users to steal their personal classified information such as login credentials, credit or debit card information, which may lead to serious monetary and identity damage for the users. One of the main challenges with this problem is the constant change in phishing URLs. Due to this, there is a constant need to update the detection mechanism, which may be extinct in a short period of time. Most of the current phishing detection tools utilise the black box method, where phishing URLs are stored and queried for verification. This may not be an efficient way due to the constant change in the URLs. In this study, a machine learning based approach is proposed along with a feature selection method to select the right set of features that may contribute to higher detection accuracy. The proposed model is also aimed at being simple, faster, and interpretable. Efficiency, accuracy, and model execution time will be evaluated against the final model.

Home Loan plays a pivotal role in today's age when one steps into purchasing their home. It has been witnessed that in many cases users are unable to pay the after taking the loan and thus the loan is slipped to NPA(Non-Performing Asset) from Standard Asset for the bank or any lending institution. The revenue generation is ceased. As the housing loan is taken against property the lenders have right to sell the property and close the dues, but the process is lengthy as judicial procedures are involved. In most cases, the property value is much less than the calculated loan amount (Principal + Interest). In this study we examined the several ML methods to identify the loan default before disbursing the loan to the applicant. This matter has been studied widely and used the predictive analytics to find out the relationship between attributes and the target variable. Predictive Analytics enables us to feed optimal set of features to the ML models. The study started with 122 attributes and ended up with around 30% of features as the ideal subset for housing loan default prediction. Then, five ML models were fit into the dataset and the champion model came up with roc score 0.94, Recall 0.90 and Precision 0.94. LIME and SHAP were applied on the champion model along with the dataset for global and local interpretability. The experimental procedure concluded that ML models along with predictive analytics can arrest the loan disbursal to the ineligible applicants and will also provide the insight of such prediction with the help of model interpretability.

In today’s job market, finding a suitable job is a complex task that requires innovation to ease the complexity of finding the job. As technology shapes the business landscape, a deeper understanding of the Advance Natural Language Processing (NLP) and using it properly will elevate the solution of finding the match between candidates and employers. In this research, comparative approach between GPT-2 and T5 model has been done to find out the best model for the job matching chatbot. Also, in this chatbot multiple criteria decision making method has been used to find the best job related to the user’s requirement. The research novelty is a chatbot that works based on the Transformer-based Text-to-Text Transfer Transformer (T5) model and compare it to GPT-2 to address the challenges of finding the best job based on job seekers’ preferences and also compare both generated answers to realise the accuracy of each model in job matching chatbot. GPT-2 and T5 both are excel in natural language understanding tasks, enabling us to parse and map user queries to many job preferences, ensuring a comprehensive understanding of user skills, and also they can provide a high rate of accuracy and performance in the natural language tasks. The research’s contribution focuses on preference job matching chatbot application, which effectively bridges the gap between employers and job seekers. Using context-based meanings of specific words and new terms defined in the conversation, the model generates responses based on user input. A seamless connection between job seekers and potential employers is made possible by our approach to Human Resources technology, which serves as a more personalised, effective, and user-friendly job matching system. The model tokenises words and generates test cases based on them. By utilising NLP techniques, it will help to ensure that all scenarios are taken into account.

In this study, we explore deep learning models for single-character Chinese character recognition tasks, with a special focus on two architectures: VGG19 and EfficientNetV2. By improving recognition accuracy with limited computational resources and dataset size, this study aims to address real-world challenges.In the field of image recognition, deep learning has shown excellent ability, especially the importance of convolutional Neural Network (CNN). This study reviews the history of handwritten Chinese character recognition and explores the application of deep learning in various fields such as object detection, image classification, and semantic segmentation. The research methodology incorporates problem analysis, data collection, model construction, training, and result generation. PyTorch is used as the basic framework to implement the model, and strict data preprocessing is conducted to optimise the performance. The user interface and interaction design allow us to show the practical application of the model and encourage user-friendly participation. By applying VGG19 and EfficientNetV2 models in a single-character Chinese character recognition task, we reveal the impact of limited training data and computational constraints on accuracy and performance. We confirm that higher training cycles improve accuracy, but we also note diminishing returns. Meanwhile, the research highlights the exciting potential of deep learning in character recognition tasks and advocates its widespread application in practice. While overcoming computational and data limitations, our study reveals the intricate relationship between model training, accuracy improvement, and practical usability. The experimental results show that the average training result of the EfficientNetV2 model is 94.957322%, and the average training result of the VGG19 model is 95.756285%. This study provides dedicated support for the in-depth research and development of Chinese character recognition and its various application fields.

Stock price forecasting has always been a challenging task due to its high volatility and complexity. Recently, various machine learning models have been employed to improve the accuracy of stock price predictions. This research aims to enhance stock price forecasting by integrating supply chain factors into LSTM models and conducting a comparative performance analysis with other models such as ANN, RNN, and GRU. A comprehensive literature review was conducted to understand the history of machine learning in finance, the role of different models in stock market prediction, the impact of varied factors in stock market prediction, and the challenges and limitations of artificial intelligence in stock forecasting. The methodology involved problem analysis, defining the research purpose, and project design which includes data collection, model building and training, and performance evaluation and validation. The models were trained and assessed on a dataset that incorporated supply chain factors. The experimental results showed that the GRU model outperformed the other models in terms of R2 score, MAE, and MSE. The study contributes to the existing body of knowledge by providing empirical evidence on the importance of incorporating supply chain factors into predictive models and by comparing the performance of different models. The findings have practical implications for investors, analysts, and policymakers who rely on accurate stock price predictions for decision-making.

With the emerge of Internet of Things, billions of devices and humans are connected directly or indirectly to the internet. This significant growth in the number of connected devices rises the needs for a new development for the current network paradigm (e.g., cloud computing). The new network paradigm, such as fog computing, along with its related edge computing paradigms, are seen as promising solutions for handling the large volume of securely-critical and delay-sensitive data that is being produced by the IoT nodes. In this paper, we give a brief overview on the IoT related computing paradigms, including their similarities and differences as well as challenges. Next, we provide a summary of the challenges and processing and storage capabilities of each network paradigm.