Office of Undergraduate Research

Permanent URI for this communityhttp://hdl.handle.net/1903/20157

Emphasizing equitable and inclusive access to research opportunities, the University of Maryland's Office of Undergraduate Research (OUR) empowers undergraduates and faculty to engage and succeed in inquiry, creative activity, and scholarship. This collection includes materials shared by undergraduate researchers during OUR events. It also encompasses materials from Undergraduate Research Day 2020, Undergraduate Research Day 2021, and Undergraduate Research Day 2022, which were organized by the Maryland Center for Undergraduate Research.

Browse

Subject: search.filters.subject.Machine Learning

Search Results

Now showing 1 - 5 of 5
  • Thumbnail Image
    Item
    Quantum-Enhanced Forecasting: Leveraging Quantum Gramian Angular Field And CNNs for Stock Return Predictions
    (2024) Bapanapalli, Abhinav; Mahmud, Ateef; Nadig, Chiraag; Thakker, Krish; Jabeen, Shabnam
    Predicting stock price movements is a complex challenge faced by many traders and analysts. Our research leverages Quantum Gramian Angular Field (QGAF) transformations combined with Convolutional Neural Networks (CNNs) to classify stock price trends as "up" or "down." By transforming 1D time-series stock data into 2D images, we enable CNNs to extract features more effectively, showcasing the potential of quantum machine learning in financial forecasting.
  • Thumbnail Image
    Item
    Attention! Data Helps Diagnoses: A machine learning approach to predicting ADHD
    (2024) Navaleza, Irene; Coutts, Jacob
    Attention deficit hyperactivity disorder (ADHD) is often dismissed as a “childhood condition”, since easy-to-identify features (e.g., hyperactivity) are more prevalent in children. Yet, for almost half of diagnosed individuals, the effects of ADHD persist through adulthood, impacting important areas such as jobs/academic performance and relationships. These implications make early diagnoses and effective treatments salient issues for medical professionals. However, as ADHD affects brain development, symptoms often greatly vary person to person. Further, research suggests that the high comorbidity of ADHD with other disorders compounds this issue, explaining why many diagnoses do not come until adulthood. One solution to more accurate diagnoses is machine learning, a class of models that have become increasingly prevalent in research. However, few researchers have developed models to predict ADHD diagnoses. In this study, we performed a secondary data analysis from a study on 103 anonymous participants (51 diagnosed with ADHD, 52 clinical controls). We employed a K-nearest neighbors algorithm to identify key features of ADHD (e.g., prevalence of comorbid disorders) that can accurately predict one’s diagnosis. The results of our analysis suggest: 1.) Objective metrics like this may improve ADHD diagnoses, since current methods are subjective and vary by physician, 2.) Some comorbidities are more predictive than others, and 3.) Research should continue in this area to include more predictive features. Implications for practitioners and researchers are discussed.
  • Thumbnail Image
    Item
    Improving Our Understanding of Dark Matter Halos with Machine Learning
    (2024) Vladimir, Ze'ev; Diemer, Benedikt
    Dark matter constitutes 80% of all matter and is key to understanding the structure of the universe. Dark matter impacts galaxy formation through the gravitational influence it exerts on baryonic matter. However, dark matter and the structures it forms are too complex to describe with an analytical solution. As a result numerical simulations have provided a way to learn about dark matter structures. Within these simulations, dark matter forms filaments, voids, and most notably halos. Halos are roughly spherical regions with a high concentration of dark matter that contain a galaxy which co-evolves alongside it. However, determining which particles form a halo is most commonly oversimplified with the spherical over-density model. This model defines a halo as all the particles present in a sphere around an over-dense region in space. However, a more physically accurate definition for halos has recently been proposed: only the particles that orbit a halo's center of mass constitute the halo. Unfortunately, current methods that determine if a particle is orbiting either lack a high level of accuracy or are computationally expensive. To fill this gap, we have developed a simple machine-learning model using the XGBoost library. Our model is trained on the velocity and position of the dark matter particles which are accessible from all simulations. To improve accuracy, we utilized velocity and position data from two snapshots in simulation time to capture how a particle moves. Our model achieves over 97\% accuracy at all radii on our validation set. In addition, our model's speed outperforms more complex orbiting classifiers on the order of seconds to minutes compared to hours.
  • Thumbnail Image
    Item
    Content Analysis Versus Structured Topic Modeling: Unethical Military Leadership
    (2021) Cheng, Virginia; Hsiao, Kimberly; Smolko, Braden; Epistola, Jordan; Hanges, Paul; Hanges, Paul
    The U.S. Military strives to reduce unethical behavior in the organization by identifying and targeting unethical leaders. This qualitative study employs content analysis and the structured topic modeling machine learning algorithm to understand unethical leadership in the Military on over 439 interviews with military personnel at six different locations. Specifically, we compared the results of both approaches to identify important aspects of unethical leadership in the Military. In general, both methods produced similar results that supported one another. Moreover, the STM method produced new insights of unethical leadership that were highly interpretable by the research team. An overview of this research project and a subset of results are presented in this poster.
  • Thumbnail Image
    Item
    Balancing Fairness and Profit in Rideshare using Deep Learning
    (2020-04-26) Raman, Naveen; Dickerson, John
    Rideshare services such as Uber and Lyft have become much more popular over the last few years. Determining which riders are allocated to which drivers is a challenging problem, complicated by the number of combinations of riders and drivers. Current solutions typically optimize for gross mean volume or profit, while ignoring fairness in driver pay and rider wait-time. We address these by taking a long-term view, using Neural Networks to simulate value functions, and use Markov Decision Processes to maximize fairness.