Data Science Commons^™

Hyperparameter Tuning For Robust Autonomous Vehicle Vision, Nico D. De Ros

Theses and Dissertations

Classification “flickering,” where the classification of an object changes inconsistently between consecutive video frames, remains a persistent issue in modern object classification algorithms. This problem undermines the reliability of autonomous vision systems and poses significant risks in high-stakes applications such as autonomous vehicles. This thesis explores the use of response surface methodology, a statistical design of experiments technique, to optimize hyperparameters across three object classification pipelines. The first pipeline combines YOLOv8 with SORT to establish a benchmark. The second integrates a Bayesian back-end, while the third employs an exponential smoothing back-end. Hyperparameter tuning was conducted using a two-step process: an …

Data Lakehouse And Machine Learning Pipeline For Aircraft Fuel Efficiency Experimentation, Skyler G. Kepley

Theses and Dissertations

Fuel efficiency is crucial for the U.S. Air Force, impacting mission success, aircraft performance, and cost savings. This study presents an information system that integrates flight and maintenance data using a data lakehouse. It automates ingestion, enrichment, and predictive modeling, leveraging AutoML for optimization and SHAP for transparency. A case study on C-130J aircraft shows that optimizing D Check cycles can save 11.52 pounds of fuel per flight hour. These findings highlight the effectiveness of data-driven decision-making in aviation, offering a scalable, automated solution for improving fuel efficiency and reducing costs.

Machine Learning With Flight Data Recorder Data For Flight Fuel Consumption Predictions, Adam C. Levandowski

Theses and Dissertations

This study applies advanced Machine Learning (ML) to Flight Data Recorder (FDR) data for fuel consumption predictions. It explores feature engineering, model selection, and Hyper-Parameter Optimization (HPO) across all flight phases. Baseline models like Ordinary Least Squares (OLS) regression, Multi- Layer Perceptrons (MLPs), and decision trees are compared to Deep Neural Networks (DNNs), Recurrent Neural Networks (RNNs) with Gated Recurrent Unit (GRU) layers, and XGBoost. Results analyze segmentation strategies, tailored features, and model performance. A counterfactual analysis compares ML models to operational fuel predictions, demonstrating their deployment potential. Findings establish a foundation for future ML-driven advancements in aviation fuel optimization.

Improving Zero Shot Learning By Linking Multi-Label Cnns With Llms, Michael A. Wegner

Theses and Dissertations

Classifying previously unseen objects poses a significant challenge for traditional computer vision algorithms, which rely on extensive labeled training data. Zero-shot reasoning offers a way to overcome this limitation. This research explores a novel method for image recognition using the Animals with Attributes 2 (AWA2) dataset as a proof of concept. A multi-label ResNet50 model predicts core attributes like color, ear shape, or number of limbs. Those attributes then feed into ChatGPT which leverages its extensive knowledge base to classify the animal based on the provided attributes. This novel approach skips the need to train on every possible class. Instead, …

Class Imbalance: A Landscape Of Classification Models, Joshua L. Edmonds

Theses and Dissertations

Class imbalance poses significant challenges in machine learning classification. This study evaluates the performance of seven models (ANN, k-Means, kNN, LDA, LR, SVM, XGBoost) across multiple imbalance levels (10\%, 5\%, 1 \%, 0.5\%) and investigates the effectiveness of sampling techniques (Undersampling, SMOTE, SMOTE-ENN). ANOVA results confirm that model choice is the most critical factor, with XGBoost and SVM demonstrating superior robustness. SMOTE improves recall but reduces precision, while undersampling generally degrades overall performance. While significant, imbalance levels do not play a critical role in model effectiveness.

A Reinforcement Learning Approach For Maneuvering And Firing Decisions In Sead Operations, Nathaniel Garcia

Theses and Dissertations

The integration of automated processes in defense continues to expand, enhancing the lethality of military forces. Artificial intelligence accelerates decision-making cycles, removes the constraints of human-operated hardware, and improves coordination by enabling seamless integration across multiple systems. Suppression of Enemy Air Defenses (SEAD) missions are critical to the United States (U.S.) military, as they neutralize hostile air defense systems, ensuring air superiority and enabling safe and effective operations for aircraft in contested environments. Therefore, it is necessary to pair emerging autonomous capabilities with an important mission set in defense. This research investigates the Autonomous Unmanned Air-to-Ground Strike (AUAGS) problem, modeling …

Proximal Policy Optimization Applied To The Beyond Visual Range Air Combat Maneuvering Problem, Daniel B. Joseph

Theses and Dissertations

Artificial intelligence (AI) grows ever-more important in warfighting. Emerging technologies allow for the use of AI to control aircraft and weapons systems. This research investigates the application of reinforcement learning (RL) through the Proximal Policy Optimization (PPO) algorithm to a two-versus-two (2v2) beyond-visual-range (BVR) air combat maneuvering problem (ACMP). Implemented in the Advanced Framework for Simulation, Integration, and Modeling (AFSIM), the methodology frames the engagement as a Markov decision process, wherein an autonomous RL agent learns continuous control decisions—throttle, pitch, roll, and yaw—under a cooperative communication scheme. A multi-phase curriculum-learning approach facilitates the progressive acquisition of flight stability, weapon deployment, …

Symbology Detection And Numerical Recognition For T-38 Heads-Up Display Recordings, Ben T. Hepner

Theses and Dissertations

The extraction of symbology and numerical data from the T-38 Heads-Up Display (HUD) enhances post-flight analysis and supports real-time decision-making. This research develops a deep learning pipeline using YOLO-based object detection and Optical Character Recognition (OCR) to analyze HUD video data. Model evaluations showed mAP0.5:0.95 ranging from 0.422 (YOLOv11m, hard test set) to 0.696 (YOLOv8m, medium test set), demonstrating robust symbology detection. Numeric detection performed well (mAP_0.5:0.95 = 0.764), but OCR struggled with glare and resolution limitations, achieving a recognition accuracy of 17.35%. These results validate deep learning for HUD data extraction but highlight the need for improved robustness …

Machine Learning Techniques To Predict Solar Particle Events And Radiation Of Aircrew, Haley Traub

Theses and Dissertations

Solar Particle Events (SPEs) are high-energy phenomena from the Sun that pose risks to technology, human health, and Air Force operations. Accurate prediction of SPEs exceeding 100 MeV is crucial for mitigating these risks. This thesis explores using Bayesian statistical models to predict such events, integrating prior knowledge from solar physics with the ability to update predictions based on new data. The research uses a dataset spanning three solar cycles (21–23) and incorporates attributes like flare fluence, peak flux, latitude, longitude, and class. Four Bayesian models (PyMC, Bnlearn, and two Dredge models) were compared to machine learning models. The Bayesian …

Topical Text Segmentation For Stream Of Consciousness Writing, Yuwei Lu, Ryan Benton

Shelby Hall Graduate Research Forum Posters

Stream of consciousness writing has a long history, including novelists James Joyce and Virginia Woolf. However, there has been little work done in automated and semi-automated analysis of such writing, which is the focus of this work. We plan to divide real streams of consciousness writing into distinct topical units and then capture different momentary meaningful topics from these units. By doing this, researchers and readers could gain a more nuanced understanding of the narrative structure and thematic elements. In addition, it would also support applications in fields like psychology and linguistics, where understanding thought processes and narrative structures is …

Comparative Evaluation Of Linear Regression, Cross Validation And Regularization Approaches In Multivariate Data Analysis, Ransford Owusu, Felix Yeboah, Francis Effah Boateng

Data Science and Data Mining

This study evaluates linear regression and its enhanced variants incorporating cross-validation and regularization techniques for high-dimensional, multivariate datasets. We address challenges such as multicollinearity and overfitting. Methods including Ridge, LASSO, and Elastic Net are compared against ordinary least squares regression. Empirical analysis using an automobile dataset for fuel efficiency prediction shows that while OLS regression captures basic relationships, its limitations are mitigated through regularization and cross-validation, resulting in improved model interpretability. The findings provide a comprehensive framework for predictive modeling in complex data environments and offer insights into statistical methodology and practical applications in the automobile industry.

Toward Quantifying Interpolation Uncertainty In Set-Line Spacing Hydrographic Surveys, Elias Adediran, Christos Kastrisios, Kim Lowell, Glen Rice, Qi Zhang

Faculty Publications

The oceans remain one of Earth’s last great unknowns, with about 74% still unmapped to modern standards. Consequently, interpolation is employed to create seamless digital bathymetric models (DBMs) from incomplete hydrographic datasets, but this introduces unquantified depth uncertainties. This study aims to estimate and characterize uncertainties arising from set-line spacing hydrographic surveys, which are important for nautical charting, navigational safety, and many other applications. By sampling at different line spacings four complete coverage testbeds that vary in slope and roughness, the study interpolates across entire testbed areas using Spline, Inverse Distance Weighting, and Linear interpolation. The resulting interpolation uncertainties are …

A Report On Health Care Access By The United States Citizens., Kelvin Njuki, Emil Agbemade

Data Science and Data Mining

Access to health care is a critical factor in ensuring public health. This study analyzes data from the National Health Interview Survey (NHIS) for the years 2015–2018 to examine the relationship between health care coverage, affordability, and costs among U.S. families. Re-sults indicate that families with at least one member covered by health insurance were more likely to afford medical care and incur lower health care costs. Despite a high proportion of families with health care coverage during this period, the number of insured family members declined over the years. These findings underscore the importance of health care coverage in …

Proxy Panels Enable Privacy-Aware Outsourcing Of Genotype Imputation, Degui Zhi, Xiaoqian Jiang, Arif Harmanci

Faculty, Staff and Student Publications

One of the major challenges in genomic data sharing is protecting participants' privacy in collaborative studies and in cases when genomic data are outsourced to perform analysis tasks, for example, genotype imputation services and federated collaborations genomic analysis. Although numerous cryptographic methods have been developed, these methods may not yet be practical for population-scale tasks in terms of computational requirements, rely on high-level expertise in security, and require each algorithm to be implemented from scratch. In this study, we focus on outsourcing of genotype imputation, a fundamental task that utilizes population-level reference panels, and develop protocols that rely on using …

Modeling The Relationship Between Calories And Activity Metrics: A Regression Analysis With Variable Selection, Felix Yeboah

Data Science and Data Mining

Physical activity monitors have become integral to daily routines, with wearable devices such as the Apple Watch and Fitbit offering continuous data on users’ physical activity. This study compares the measurement accuracy of these devices by examining how they record parameters relevant to fitness and health. Employing multiple linear regression, we modeled the relationship between calories expended and a set of explanatory variables, including heart rate, steps, distance, age, activity level, weight, and device type. Evaluation of all possible variable combinations identified heart rate, steps, distance, weight, and watch type as the most effective predictors of calorie expenditure. Although the …

Locational Data And The Public Interest, William A. Herbert, Micahel Goodchild, Richard Appelbaum, Jeremy Crampton, Gary Langham, Krzysztof Janowicz, Mei-Po Kwan, Katina Michael, Lisa Schamess

Publications and Research

This article presents a paper developed by the AAG Organizing Committee on Locational Information and the Public Interest through a summit held in Santa Barbara, California in June 2022. The summit resulted in goals and ideas for addressing the issues that arise from the present environment for geodata, whereby public, private, and third-sector entities can tap into publicly available locational information with relatively little regulation on its access or use. The Committee articulates four goals: (1) develop a research agenda extending across disciplines, (2) outline educational resources and strategies to guide ethical practice, (3) devise a pathway to increase public …

Bayesian Machine Learning Approach For Corn Yield Prediction Using Satellite Imagery And Topographic Data, Etornam Kwame Kunu, Hossein Moradi Rekabdarkolaee

SDSU Data Science Symposium

In an era of climate change and growing global food demand, accurate crop yield prediction is pivotal for leveraging advanced technologies to enhance crop management and sustainability. This study compares the prediction performance of several Bayesian Machine Learning method using high-resolution PlanetScope imagery and topographic data. In specific, the Bayesian Linear Regression, Bayesian Random Forest, Bayesian Splines, Bayesian Additive Regression Trees, and Bayesian Neural Network were developed to incorporate uncertainty quantification and achieve enhanced predictive accuracy. Our finding shows that the Bayesian Random Forest outperform the other model in term of crop yield prediction.

Classification Of Schizophrenia, Bipolar Disorder And Major Depressive Disorder With Comorbid Traits And Deep Learning Algorithms, Xiangning Chen, Yimei Lu, Joan Manuel Cue, Mira V Han, Vishwajit L Nimgaonkar, Daniel R Weinberger, Shizhong Han, Zhongming Zhao, Jingchun Chen

Faculty, Staff and Student Publications

Many psychiatric disorders share genetic liabilities, but whether these shared liabilities can be utilized to classify and differentiate psychiatric disorders remains unclear. In this study, we use polygenic risk scores (PRSs) of 42 traits comorbid with schizophrenia (SCZ), bipolar disorder (BIP), and major depressive disorder (MDD) to evaluate their utilities. We found that combining target specific PRS with PRSs of comorbid traits can improve the classification of the target disorders. Importantly, without inclusion of PRSs from targeted disorders, we can still classify SCZ (accuracy 0.710 ± 0.008, AUC 0.789 ± 0.011), BIP (accuracy 0.782 ± 0.006, AUC 0.852 ± 0.004), …

Optimized Hiv/Aids Resource Allocation In Ohio: A Linear Programming Approach, Godfred Ahenkroa Kesse

Data Science and Data Mining

This study employs a linear and integer programming approach to optimize HIV resource allocation in Ohio, aiming to minimize new infections and enhance the impact of limited resources. With the advances in HIV prevention and treatment, Ohio faces challenges in addressing disparities in access to healthcare, particularly among high-risk populations. The proposed model integrates data on infection rates, transmission patterns, demographic factors, and cost-effectiveness to provide a decision-support framework for policymakers. Using epidemiological data and equity constraints, the model prioritizes high-risk regions and populations while ensuring fair resource distribution. Results indicate that increased funding allocations significantly enhance the potential to …

Assessing Vegetation Changes And Ecosystem Impacts Of Invasive Species In Florida's Water Conservation Area 3a (2004–2024) Using Landsat Data, Gideon Tandoh

Research, Papers & Creative Work

The spread of invasive species is a concrete problem, threatening biodiversity, ecosystem integrity, and the functional status of wetlands. As these species continue to advance, it is crucial to assess their spread as well as management success through the more sophisticated invasive monitoring methodologies. This research examines the change in vegetation communities and the spread of invasive plants in the Water Conservation Area 3A (WCA-3A) using Object-Based Image Analysis (OBIA) and Support Vector Machine (SVM) classifiers on time series Landsat images. Unlike traditional pixel-based methods, OBIA enhances classification accuracy by incorporating spatial, spectral, and contextual information, enabling precise differentiation between …

Testing Toxicity: Two Approaches To Toxicity Detection, Nelson E. Jarrin

Dissertations, Theses, and Capstone Projects

The current online environment is increasingly being polluted by polarization, echo chambers, and a lack of constructive dialogue between individuals with differing viewpoints. Social media algorithms, while intended to personalize user experiences and marketing areas, have been heavily criticized for inadvertently amplifying extremist content and contributing to the spread of toxic commentary. The current state of the internet hinders productive conversations, fuels misinformation, entices disinformation and contributes to online and offline societal division. This is fueled by algorithms that can create echo chambers and by the anonymity given by online platforms. While tools like the Perspective API exist to identify …

Quantified Lives: Data, Bias, And The Cost Of Categorization, Colin F. Geraghty

Dissertations, Theses, and Capstone Projects

This project critically examines how statistical methods and data visualization have historically been used to categorize, rank, and control human populations. By exploring the works of key figures like Francis Galton, Karl Pearson, and R.A. Fisher, it traces the evolution of categorization frameworks from their eugenic roots to modern applications in artificial intelligence and global metrics like the World Happiness Report. Through interactive visualizations and historical analysis, the project invites users to question the legitimacy of these frameworks and reflect on their impact on society. At its core, this work emphasizes the need to move beyond rigid classification systems, challenging …

Heartdj - Music Recommendation And Generation Through Biofeedback From Heart Rate Variability, Egemen Şahin

Dartmouth College Master’s Theses

This study investigates the integration of real-time physiological data with AI-generated music to enhance emotional well-being, stress regulation, and focus, using Heart Rate Variability (HRV) as a biomarker of autonomic function. Conducted in two phases—Stable Audio Open (SAO) and Suno (SUNO)—the research evaluates biofeedback-driven music interventions across varying daily music-listening habits.

In the SAO phase, short AI-generated instrumental tracks were compared with Spotify recommendations and guided meditation. Modest HRV improvements were observed in biofeedback conditions, but participants noted emotional limitations, citing short track lengths and abrupt transitions.

The SUNO phase addressed these limitations with longer, more complex AI-generated compositions combined …

In Memoriam - Nora Sabelli: Master Orchestrator Of Grant Programs And Mentor For Advancing The Interdisciplinary Learning Sciences Field, Eric Hamilton, Jeremy Roschelle, Roy Pea, Barbara Means, Louis Gomez, Kim Gomez, Nancy Butler Songer

Education Division Scholarship

On Friday, September 6, 2024, the learning sciences field lost a giant in Dr. Nora Sabelli, 87 years old, a personal mentor to many researchers and an inspiration to so many learning scientists and STEM leaders. Nora’s first professional career was as a computational chemist, and later she became a passionate leader in research for improving STEM education. Nora’s time as a senior program officer at the National Science Foundation’s (NSF) Education and Human Resources (EHR) directorate was legendary; she was a force of nature who reshaped funding priorities for stronger science and a stronger connection of science to education …

Uncovering Acoustic Biomarkers To Classify Parkinson Disease Through Machine Learning, Felix Yeboah

Data Science and Data Mining

The early detection of diseases profoundly influences treatment efficacy, and accurate classification methodologies are essential for effective disease identification. In this project, we examined fve different classifers—Logistic Regression, Gaussian Naive Bayes, K Nearest Neighbor (KNN), Extreme Gradient Boosting (XGBoost), and Support Vector Machines—and evaluated their performance in detecting Parkinson’s disease (PD) based on voice features. The study aims to identify the best classifier for detecting PD. XGBoost performed the best, with an accuracy of 91% on the full dataset. After variable selection, KNN had the best performance with an accuracy of 91%. These findings suggest that Machine learning algorithms(classifiers) can …

Evaluation Of Variable Selection Techniques On The Genetic Architecture Of Flowering Time In Maize, Felix Yeboah

Data Science and Data Mining

In this project, we investigate several variable selection procedures to give an overview of how well they perform on a genomic dataset using three different penalized regression approaches. Comparisons between different methods were performed. These methods include Ridge, lasso, and Elastic Net. We utilized 4494 observations with 7389 SNPs gene scores to predict time to male flowering (dtoa). We assessed the performance of these three models in terms of mean square error. Not surprisingly, Lasso and Elastic Net perform better than Ridge Regression. Overall, Elastic Net performed better in predicting the time of male flowering (dtoa).

Comparison Of Two Strategies Of Screening Experiments: Single-Shot Experiment Vs. Two-Stage Screening Experiment, Kelvin Njuki, Emil Agbemade

Data Science and Data Mining

Experiments involving many factors are often complex, time-consuming, and expensive. Screening out the least important factors helps the experimenter(s) allocate the limited resources efciently to the most important factors. Supersaturated and orthogonal array designs are among the designs used to conduct screening experiments. Supersaturated designs (SSDs) are those where the number of runs (observations) is less than the number of factors, while orthogonal array (OA) designs are those where at least the columns are orthogonal to each other. In this study, we conduct a simulation study to compare two strategies of screening experiments. Strategy one is a single shot experiment …

Pull Or Play? The Interpretation Of A Novel, Ai-Powered On-Field Decision Support Tool., Lynne Becker, Dafne Badilla, Osho Yonzon, Xiaoyu Wu, Roland Rocafort, Erik Viogan Phd, Devansh Manocha

Journal for Sports Neuroscience

The document titled "Pull or Play? The interpretation of a novel, AI-powered on-field decision support tool" explores the development and application of the Injury Impact Severity Score (IISS)™ for assessing traumatic brain injuries (TBIs), particularly in sports settings. It addresses the limitations of current assessment tools like the Glasgow Coma Scale (GCS) and proposes a more objective approach using patient-reported data and machine learning algorithms.

Key points include:

• Background: TBIs are a significant health concern with under-reported cases and a lack of effective research. Current assessment tools like the GCS have limitations in accuracy and speed, especially in dynamic …

Advanced Machine Learning Techniques For Cardiovascular Disease Risk Prediction, Godfred Ahenkroa Kesse

Data Science and Data Mining

of mortality, necessitating advanced predictive models to aid early detection and prevention. This study explores the application of machine learning techniques, including Lo- gistic Regression, K-Nearest Neighbors (KNN), Random Forest, and XGBoost, to predict CVD risk using a dataset of 69,997 observations encompassing demographic, clinical, and lifestyle factors. Data preprocessing involved one-hot encoding of cat- egorical variables and scaling to ensure compatibility with all models. Model performance was evaluated using metrics such as accuracy, precision, recall, F1-score, and AUC-ROC. Among the models, XGBoost demonstrated the highest accuracy at 74%, leveraging its gradient-boosting framework to effectively handle feature interactions and imbalanced …

Predicting Blood Glucose Levels: A Linear Regression Approach For Non-Invasive Monitoring, Godfred Ahenkroa Kesse

Data Science and Data Mining

Accurate monitoring of blood glucose levels is vital for the management of diabetes, a chronic condition affecting millions worldwide. This study explores a linear regression approach to estimate glucose levels non-invasively using a dataset enriched with demographic, physiological, and sensor-based variables. Following rigorous data preparation, including normalization and encoding, a Box-Cox transformation was applied to address violations of regression assumptions, stabilizing variance and improving model validity. Stepwise selection and hypothesis testing were employed to refne the model, retaining signifcant predictors such as AGE, GENDER, HEARTRATE, and DIABETIC, while excluding variables like NIR Reading and LAST EATEN for their minimal contribution. …