Research
Three research threads, kept deliberately separate.
Explainable AI applied to student outcomes, interpretable deep learning, and an in-progress behavioral prediction study — different scopes, different stages of completion, presented here exactly as far as each has actually gone.
Beyond Prediction: Benchmarking ML, XAI, and AI-Generated Interpretations of Student Academic Outcomes
My current major research project, sitting at the intersection of machine learning, explainable AI (XAI), and large language models as applied to interpreting student academic outcomes. The work benchmarks predictive machine-learning approaches, investigates explanation methods including SHAP and LIME, and evaluates how AI models interpret those predictions and explanations in turn.
This project is actively in progress. No results, findings, or conclusions are reported here yet — this page will be updated as the work develops toward its September 2026 AI Symposium presentation.
Project shape
01 — Benchmark
Compare predictive ML approaches for modeling student academic outcomes.
02 — Explain
Apply explanation methods, including SHAP and LIME, to those models' predictions.
03 — Interpret
Evaluate how AI models themselves interpret those predictions and explanations.
Related Research Notes
Advanced Deep Learning Research Assistant
Supervisor: Dr. Wang Longwei
This role focused on prototype-based, interpretable deep learning — a family of image-classification approaches designed so a model's reasoning can actually be inspected, rather than treated as a black box. I studied three related systems in depth:
- — ProtoPNet — the original this-looks-like-that prototype architecture
- — TesNet — a transparent, evidence-based extension of the prototype idea
- — ProtoPool — a model using differentiable, shared prototype assignment
My work included:
- — Studying the research papers, model architectures, and source code behind each system
- — Understanding their training pipelines in PyTorch
- — Writing technical research reports for my supervisor on prototype-based and explainable/interpretable AI, including differentiable prototype assignment
- — Reproducing and training the ProtoPool model, following the project's public GitHub repository with slight modifications
This was a research-assistant position, not an original publication. I reproduced and trained the existing ProtoPool model — I didn't create it, and this work hasn't been presented at a conference.
How the three systems relate
01 — ProtoPNet
Learns a fixed set of class prototypes and classifies by comparing an image's patches to them.
02 — TesNet
Reworks the prototype space to be more transparent and better separated between classes.
03 — ProtoPool
Shares a pool of prototypes across classes using a differentiable assignment — fewer prototypes, less manual tuning.
Using Machine Learning to Predict Habit Formation and Maintenance in College Students
Supervisor: Dr. Isaiah Cohen · University of South Dakota, Department of Anthropology and Sociology
This project asks whether behavioral, psychological, and contextual signals — sleep, mood, stress, energy, and daily routine — can be used to predict whether a college student will successfully form and maintain a habit. It's a genuinely in-progress project, presented here at exactly the stage it's at.
Current and planned components:
- — Literature review and research planning
- — Survey development covering habit adherence and behavioral signals
- — Candidate models: Logistic Regression, Decision Tree, Random Forest
- — Candidate features: habit adherence, mood, stress, energy
No sample size, accuracy figures, or statistical results are reported here because none exist yet — the broader project remains a work in progress.
