22505 research outputs found
Sort by
Meshed Trees: A Framework for Resilient Network Algorithms and Protocols
No full textThe continued growth of computer networks in both size and complexity presents challenges for the control protocols that are tasked with maintaining optimal data-forwarding paths. These protocols commonly rely on shortest-path algorithms to determine where network data should be propagated and how it should be analyzed for inclusion in forwarding tables. However, problems arise during reconvergence, the process in which devices rediscover optimal paths following a network failure. Because control protocols were built decades ago for the networks of that era, the existing shortest-path algorithms utilized were designed for problems before the advent of modern computer networks. This resulted in protocols that often cause extended downtime while reconverging the network to a functional state. This downtime reduces service availability, reliability, and can result in a substantial financial impact. This is especially true for data centers, which house thousands of servers to provide cloud-native services. This dissertation addresses these challenges through the development of proactive solutions to minimize downtime and reduce the computational overhead required for reconvergence. Proactivity is achieved through the construction of a Meshed Tree, a novel structure designed for resiliency by maintaining both preferred and precomputed backup forwarding paths. These backup paths, managed by the Meshed Tree Algorithm (MTA) introduced in this work, mitigate reconvergence delays when a component failure occurs along the preferred path. MTA was validated and implemented in a Meshed Tree Protocol (MTP) for use in switched Ethernet networks and then extended for data center networks through the use of a multi-rooted Meshed Tree. By explicitly designing an algorithm that accounts for the dynamic nature of modern networks, MTPs can significantly reduce latency, control overhead, and data loss during recovery. Three solutions implementing Meshed Trees are presented in this dissertation: 1. The Meshed Tree Algorithm (MTA): A distributed algorithm that stores and ranks path-vectors at every node to determine their preferred and backup paths. These path-vectors are used to build both the shortest-path tree and alternative trees to be used in the event of a failure. An analysis of MTA is presented and then compared to Dijkstra’s algorithm and the IEEE Spanning Tree algorithm to prove its proactive approach allows for an efficient reconvergence system. 2. Meshed Tree Protocol for Switched Networks (MTP-SW): A protocol enhancing network-wide resiliency in switched Ethernet environments by introducing a Meshed Tree. MTP-SW extends prior evaluations against the Rapid Spanning Tree Protocol (RSTP) to examine root node failures and to demonstrate how a decoupled forwarding system for unicast traffic enables efficient path reconstruction. 3. Meshed Tree Protocol for Data Center Networks (MTP-DCN): A protocol integrating the Meshed Tree system into folded-Clos topologies. By designing a control protocol around recurring topological patterns, MTP-DCN achieves a marked reduction in operational complexity. Its performance is evaluated against the Border Gateway Protocol (BGP), which is regularly used in DCNs, to demonstrate improvements in reconvergence time, packet loss, and control overhead. These contributions demonstrate that by rethinking how routing and forwarding are approached in structured network topologies, it is possible to achieve faster recovery, reduced control overhead, and improved overall resiliency
EARLY DIABETES PREDICTION USING MACHINE LEARNING: A COMPARATIVE STUDY OF CLASSIFICATION MODELS
No full textThe current dissertation investigates the application of machine learning in diabetes prediction at early stages through a comparison of the performances of three classification models, including Logistic Regression, Decision Tree, and Random Forest. Inspired by the increased and spread cases of diabetes worldwide, the research objective is to facilitate early diagnosis by providing interpretable and accurate predictive models. Based on Pima Indians Diabetes Dataset containing 768 clinical records, the research implemented data preprocessing including KNN imputation and outlier processing, feature scaling, and formation of interaction features. Quantitative and comparative approach was made to train and test the models based on such metrics as accuracy, precision, recall, F1 score, and ROC-AUC. The Logistic Regression proved to have the optimal level of accuracy and interpretability (AUC 0.88), that is reasonable to be used in clinical practice. Random Forest showed the best predictive accuracy although it was not transparent. The most influential features identified in prediction were glucose, BMI and age. The research suggests the usage of Logistic Regression as the solution deployed in mobile and community-based screening devices. The future research should involve more significant and heterogeneous samples of data and implement interpretability techniques like SHAP or LIME and expand the framework to predict some other chronic conditions, including hypertension
Exploring the Perception of Dynamic Electrotactile Patterns on the Fingertips
No full textThe goal of this study is to evaluate the effectiveness of electrotactile feedback on the fingertips by examining how accurately different dynamic patterns can be perceived. We designed three pattern types (Single Line, Double Line, and Box) and programmed each to move in eight distinct directions: bottom to top, top to bottom, left to right, right to left, bottom-left to top-right, top-right to bottom-left, bottom-right to top- left, and top-left to bottom-right. A study was conducted to assess these patterns’ perceived accuracy and response time on fingertips. The study results revealed that the Single Line and Double Line patterns had higher recognition accuracy than the Box pattern. Participant feedback on the Box pattern was mixed, with some finding it useful while others struggled with tracking its movement. In contrast, Single and Double Line patterns received positive feedback. Based on these findings, a second study was conducted for walking conditions. In this second phase, the participants again showed higher accuracy and stronger preference for the Double Line pattern. Hence, we conclude that the Double Line pattern provides clearer and more reliable electrotactile feedback, even during movement
DeFaking Deepfakes: Designing and Evaluating AI-Powered Digital Media
No full textThe rapid spread of generative AI has revolutionized media production, creating new challenges for information integrity as convincing deepfakes proliferate. Journalists play a critical role in upholding credible public information, yet existing deepfake detection technologies often overlook their specific workflows and requirements. This dissertation addresses these needs by identifying what journalists require from detection tools and evaluating usability in realistic scenarios through the following works: Journalists\u27 Needs and Tool Design: Through qualitative user studies, we uncover journalists\u27 preferences for tools that provide transparent, explainable evidence and context-aware analysis integrated into news verification routines. These insights drive the design of DeFake, a detection tool that integrates intra-frame, inter-frame, audio manipulation, and speaker identity analysis to support nuanced decision-making. Scenario-Based Evaluation: We present Dungeons & Deepfakes (D&DF), an innovative scenario-based role-play methodology for testing detection tools with journalists in the United States and Bangladesh. The evaluation demonstrates how time pressure, story impact, and explicit suspicions shape tool usage and reveals the prevalence of automation and confirmation biases. Simulation-Based Learning for Critical Media Verification: Building on insights from earlier phases, we design and examine a simulation-based learning environment that leverages AI-generated scenarios to support critical verification practice. The platform extends the D&DF framework into an interactive educational prototype, enabling users to engage in structured verification simulations. A small pilot study evaluates the system’s feasibility, usability, and pedagogical potential, showing that the approach is workable and offers promising directions while highlighting necessary refinements for future educational deployment. Collectively, these contributions advance practical solutions for media verification and empower journalists to safeguard information integrity within an AI-driven landscape
Three-dimensional CFD analyses of natural circulation in a pressurized water reactor during a loss of coolant accident
No full textIn this research, ANSYS FLUENT simulations will be performed on natural circulation cooling performance during loss of coolant accident on the semi-scale Mod-2A geometry. Natural circulation experiments were performed in the semi-scale Mod-2A test facility which is a small-scale model of the primary system of a four-loop PWR nuclear power generating plant. The CFD results are validated using the tests done in the facility. The scope of the research is to understand the flow, temperature distribution, inventory coolant losses, and pressure drop in the system. These findings can provide insights into the performance of such systems and identify potential improvements to enhance their reliability and efficiency. Also, CFD contours can further help to understand the flow characteristics and properties at any point of interest in the system
Predicting Bus Ridership in Dubai Using Advanced Machine Learning Models
No full textIn rapidly urbanizing cities, the public transport systems have to be efficient and reliable with proper demand forecasting to guarantee customer satisfaction. This paper explores how evidence-based machine learning tools can be used to forecast weekly bus ridership in Dubai, using a dataset of more than 10 million trip-level boarding observations between June 2013 and July 2014. It was aggregated to 25,195 route-week observations, and preprocessed with temporal feature engineering, lag variables and rolling statistics. Three predictive models have been tested: Linear Regression (as a baseline), Random Forest and XGBoost (as an ensemble method). Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Coefficient of Determination (R²) and Mean Absolute Percentage Error ( MAPE ) were used to evaluate model performance. Findings indicate that the ensemble models outperformed the baseline in all cases, and the Random Forest provided the lowest MAE and XGBoost provided the most competitive values of MAPE. The explanatory power of models showed that short-term lags and rolling averages prevailed in the formation of ridership forecasts by feature importance analysis and SHAP values. The results indicate that machine learning models can be valuable decision-support solutions to the Roads and Transport Authority in Dubai, which can be used to plan routes in advance, allocate resources efficiently, and deliver services more effectively. The paper ends with recommendations on how to expand the data coverage, combine exogenous variables and consider more sophisticated modelling methods to increase the predictive power even more
Physics Meets Data: Merging Physics-Based Methods with Deep Learning to Model Complex Systems
No full textAccurate modeling of complex systems is crucial in domains such as healthcare, where personalized diagnosis and treatment planning are essential. Traditional physics-based models provide structured, theoretically grounded insights but are often computationally intensive and constrained by simplified assumptions that limit adaptability to patient-specific conditions. In contrast, data-driven models are computationally efficient and capable of capturing complex patterns, yet they often lack interpretability and fail to incorporate essential physical principles, reducing robustness and generalization. This disconnect between mechanistic understanding and computational practicality presents significant challenges in critical applications such as healthcare, where both physical accuracy and real-time performance are vital. To address these challenges, this dissertation presents hybrid modeling frameworks that integrate the predictive power of data-driven approaches with the theoretical rigor of physics-based modeling, aiming to develop accurate and interpretable models for real-time applications in complex systems. It advances these goals through two primary research directions: (1) residual modeling, which bridges gaps between physics-based predictions and empirical data, and (2) physics-guided learning, which embeds physical laws within neural network training. The first research direction develops a residual modeling framework that corrects systematic errors in physics-based models and reconstructs hidden dynamics in complex systems. By treating discrepancies between observed data and model predictions as learnable residuals, this approach enhances predictive accuracy and captures low-amplitude signals masked by dominant patterns. The framework is demonstrated in Electrocardiographic Imaging (ECGI), where it enables accurate cardiac modeling through both error correction and dynamic reconstruction. The second research direction advances Physics-Informed Neural Networks (PINNs) through meta-learning to enable rapid adaptation across diverse configurations. First, a Difficulty-Aware Task Sampler (DATS) is introduced to improve training efficiency by prioritizing tasks based on complexity, enabling PINNs to generalize effectively while reducing computational cost. Second, a meta-PINN framework for cardiac modeling is developed that leverages hypernetworks to generate patient-specific models, enabling fast personalization without retraining. Collectively, these contributions advance hybrid modeling by unifying physics-based and data-driven approaches, resulting in robust, interpretable models with improved computational efficiency. Extensive evaluations on simulated and real-world datasets demonstrate significant improvements in accuracy, adaptability, and computational efficiency over state-of-the-art approaches, strengthening the foundation of hybrid modeling and enabling interpretable, subject-specific models for complex, real-time applications in healthcare and beyond
Persuasive Pathways in Digital Apologies: The Role of Persuasion Routes in Engagement with YouTuber Apology Videos
No full textThis study explores how viewers process YouTube apology videos and how those processing methods relate to their judgments of the creator and their willingness to continue engaging with the channel. Guided by the Elaboration Likelihood Model, a single-session online survey experiment randomly assigned 413 adult participants to watch one of five publicly available apology videos from well-known YouTubers. After viewing, participants reported the extent to which they engaged in central versus peripheral processing, rated the creator\u27s credibility and the apology\u27s sincerity, and indicated their intentions to engage. Central processing was positively associated with engagement intentions, whereas peripheral processing was weakly negatively associated. Credibility and sincerity were positively associated with engagement intentions, whereas peripheral processing was associated with lower credibility and sincerity. Comparisons between unfamiliar and familiar viewers indicated that relationships among processing, appraisals, and engagement were more pronounced among unfamiliar viewers; among familiar viewers, engagement intentions were more closely related to credibility and sincerity than to processing measures. Overall, the findings highlight central-route processing and perceived credibility as key predictors of intended engagement in digital apology contexts and suggest that peripheral cues and viewer familiarity condition responses to YouTuber apologies
AI-POWERED MOBILE PHONE ACTIVITY INSIGHTS: DEVELOPING PREDICTIVE MODELS FOR SMARTER DECISION-MAKING
No full textThis study investigates how artificial intelligence can enhance telecom network management by forecasting internet usage, predicting congestion, and identifying user behavior patterns from mobile phone activity data. The study made use of anonymized logs for calls, SMS and internet, and put up a multi-model analytical pipeline, which was composed of time-series forecasting (ARIMA, LSTM), clustering (K-Means), and classification (XGBoost), to perform the analysis. Among the time-series methods, ARIMA ranked first in the forecast performance (RMSE=0.31) and gave LSTM a convincing defeat in the case of this particular short and stable dataset. Based on K-Means segmentation, users were sorted into five behavioral groups according to their communication patterns, and each group had its own distinct profile of usage. In the case of the XGBoost classifier, when it was tested for predicting periods of high congestion, it scored 90
Surveillance Report: Che-Wei Hsu
No full textFace ID: A Surveil of Che-Wei Hsu is a lens based artistic investigation and experience into the nature of surveillance in modern society. Through performance, photography, and installation, my work explores the dynamics between being watched and watching, questioning how surveillance influences human behavior and identity. In my art practice, I think about the behavior that people would not do in front of other people or violate the social norm and perform in the public as a way to experience how surveillance influences how we act in the public, for example eating a sandwich on the ladder in front of the webcam. Inspired by the ancient penalty in China and Japan, which the government would tattoo on criminals’ bodies or faces, therefore these people would be isolated by the public. In one of my performances, I let people doodle on my face as the metaphor of how people surveil and isolate each other. Through performance, face painting, and absurd interactions with surveillance cameras, I reflect on the paradox of concealment and exposure. My practice ultimately serves as a critique of surveillance culture, highlighting its contradictions and ambiguities while inviting the audience to reconsider their role within these power structures