SELECTED WORKS · VOL. I · 2021–2026
Abstract: I teach machines to read the structure hidden in sequences. My research turns video, electricity grids, glucose traces, and proteins into symbolic graphs: de Bruijn graphs, borrowed from genomics, so that neural models can reason over them. I am a doctoral researcher in Computer Science at Indiana University Bloomington, with eleven works spanning Scientific Reports, IEEE Access, Machine Learning: Science and Technology, and an ICML 2025 workshop.
Index Terms: de Bruijn graphs, video understanding, time-series forecasting, self-supervised learning, protein classification
Mert Onur Cakiroglu is a doctoral researcher in Computer Science at Indiana University Bloomington's Luddy School of Informatics, Computing, and Engineering, advised by Prof. Mehmet M. Dalkilic and co-advised by Dr. Hasan Kurban. His work asks a single question in many domains: what does a neural model gain when a continuous signal is first re-written as a symbolic graph?
The answer, so far, spans an unusual range for an early-career researcher: forecasting electricity demand across power grids, anticipating hypoglycemic events in children with Type 1 diabetes (published in Scientific Reports with clinicians at Sidra Medicine), classifying proteins from sequence alone, and, most recently, judging whether AI-generated video moves the way the real world does. Before the doctorate he was a research associate at Texas A&M University at Qatar, building self-supervised and federated learning systems for video in the compressed domain, and a full-stack engineer shipping fault-management systems for national telecom infrastructure.
He is first author on ten of the eleven works catalogued below.
PLATE I · TIME SERIES
A frequency spectrum tells a forecaster what repeats, and nothing about why. This work maps out when contextual information genuinely improves a time-series model, and when the spectrum alone was always going to be enough.
PLATE II · VIDEO
Do deepfake detectors generalize, or merely memorize their training set? A six-control audit protocol that tells the difference, released as an open toolkit, leaderboard, and model zoo.
arXiv:2607.13006, 2026 [PREPRINT]
arXiv:2606.31004, 2026 [PREPRINT]
arXiv:2602.16086, 2026 [PREPRINT]
arXiv:2511.13897, 2025 [PREPRINT · IN SUBMISSION]
IEEE Open Access Journal of Power and Energy, 12, 691–701, 2025 [JOURNAL]
ICML 2025, 1st Workshop on Foundation Models for Structured Data, Vancouver, 2025 [ICML WORKSHOP]
IEEE Access, 13, 123182–123198, 2025 [JOURNAL]
International Symposium on Signals, Circuits and Systems (ISSCS), Iasi, Romania, pp. 1–4, 2025 [CONFERENCE]
Scientific Reports, 14, 31251, 2024 [NATURE PORTFOLIO] [CLINICAL COLLABORATION]
Machine Learning: Science and Technology, 5(3), 035020, 2024 [IF 6.8] [OPEN ACCESS]
IEEE International Conference on Data Science and Advanced Analytics (DSAA), San Diego, pp. 1–3, 2024 [EXTENDED ABSTRACT]
I.
Most video models decode every frame and start from pixels. The compressed stream already contains a physics report: motion vectors, block by block. This thread builds evaluation and detection methods that read it directly, from temporal-realism scoring of generated video to generalization audits for deepfake detectors, with self-supervised and federated learning that never leaves the compressed domain.
REFS [2] [4]
II.
Discretize a continuous signal and it becomes a string; string algorithms then apply. The DRAGON encoder maps multivariate series onto de Bruijn graphs, recovering long-range context through graph attention and improving forecasts of electricity demand, long-horizon grid load, and beyond.
REFS [1] [5] [6] [7] [8]
III.
The same graphs that assemble genomes can watch a glucose trace. With clinicians at Sidra Medicine, this thread forecast pediatric hypoglycemia with reinforcement learning over extended de Bruijn graphs, and engineered protein-classification features from sequence structure alone.
REFS [9] [10]
Before the doctorate: two years as a full-stack engineer at Innova IT Solutions, building MARS, a centralized fault-management system for national telecommunication networks with end-to-end detection, diagnosis, and resolution at infrastructure scale. Spring Boot microservices over PL/SQL and MongoDB; React and Vaadin front ends; production deadlines. The research benefits daily: models ship as software, and software is a sequence of decisions too.
Off the record, the author plays electric guitar: one more instrument that turns a continuous signal into something with structure. Six strings, one pickup, and a time series you can actually hear.