Trustworthy evaluation
Designing tests that reflect future deployment rather than random data splits.
Welcome to my homepage
Pengyu Feng
Undergraduate researcher in reliable machine learning and graph representation learning.
The Ohio State University · Computer Science and Engineering · Sophomore
I’m Pengyu. I enjoy turning complex real-world questions into testable machine-learning experiments, with current interests in reliable graph learning, temporal distribution shift, and high-stakes applications.
Research focus
Questions I am working on
Graph representation learning
Preserving local node-level signals while learning from network structure.
Applied AI
Studying machine learning where false positives, drift, and limited labels matter.
Selected publications
Research and publications
Each entry links to a dedicated page with the abstract, protocol, reported results, and representative figures.
2026
Accepted full paper · TELEPE 2026
Exploring Topological Resilience: A Decoupled Dual-Pathway Graph Neural Network for Bitcoin Anti-Money Laundering under Concept Drift
A strict out-of-time evaluation of Bitcoin AML models using only 94 local transaction features. The paper introduces TSG-Net, a dual-pathway graph model that separates topology learning from local-feature preservation.
Paper-reported test result: 0.96 illicit precision, 0.61 recall, and 0.75 F1.
2025
Published · IEEE ACCTCS 2025
Application of Machine Learning and Deep Learning in Medical Image Segmentation—an Example of Liver CT Images
A comparative study of conventional machine-learning and convolutional approaches to liver CT segmentation, covering preprocessing, evaluation, and potential improvements to U-Net.
DOI: 10.1109/ACCTCS66275.2025.00026
Education & research
Experience
The Ohio State University
B.S. student, Computer Science and Engineering · Sophomore
Current interests include trustworthy machine learning, graph learning, temporal distribution shift, and financial forensics.
Temporal graph learning for Bitcoin AML
Independent research · Accepted full paper, TELEPE 2026
Built a leakage-aware evaluation protocol and investigated a dual-pathway GNN under a strict chronological split.
Liver CT image segmentation
Independent research · Published in IEEE conference proceedings
Completed my first full research cycle, from literature review and model comparison to analysis and academic writing.
About
Researching the gap between benchmark performance and real use.
I began with medical image segmentation in my freshman year. That project taught me how to structure a research question, compare methods, and carry a manuscript through publication.
My current work moves toward stricter evaluation: chronological data splits, information-limited settings, class imbalance, and transparent reporting of failure cases. I am especially interested in research collaborations involving reliable graph learning and high-stakes AI.
Contact
Research conversations and collaboration
The best way to reach me is by email.