A high school student can do publishable data science research by starting with a question that matters in some field, finding a public dataset that can answer it, and analyzing it with methods rigorous enough to survive peer review. What gets a student data science paper accepted is rarely a fancy model. It is a clear question, an honest evaluation, and a result that holds up when a reviewer checks for the usual mistakes.
This guide covers what data science research looks like, example questions, where to find data, the rigor checks reviewers look for, where student data science papers get published, which competitions fit, and the papers YRI Fellows have written. For a coding-first view, see our guides to computer science research and AI and machine learning research.
What makes data science research different
Computer science research often asks how to build a better algorithm or system. Data science research asks what the data tells us about the world, and how sure we can be. It sits between statistics, computing and a domain such as medicine, economics, transport or astronomy.
| A data project | A data science research paper |
|---|---|
| "I trained a model on a Kaggle dataset" | "Does the model catch the high-risk cases, or only the easy ones?" |
| Reports one accuracy number | Reports several metrics, with uncertainty |
| Random train and test split | Split that matches how the model would be used |
| No comparison | Compared against a baseline or prior work |
| "It works" | "It works here, fails there, and this is why" |
That right-hand column is where the publishable contribution lives.
Areas of data science research for students
| Area | Example focus | Typical data |
|---|---|---|
| Health and medicine | Screening, risk prediction, literature mining | Public clinical and survey datasets |
| Systems and operations | Anomaly detection in servers, networks, satellites | Telemetry and event logs |
| Transport and cities | Traffic, emissions, incidents | Camera images, open city data |
| Business and economics | Who responds to an offer, credit risk | Public experiment and loan data |
| Science | Genomics, astronomy, climate | Public scientific archives |
| Text and language | Classifying documents or research findings | Text corpora, literature databases |
Example research questions
- Does a model that predicts disease risk still catch high-risk patients when they are rare in the data?
- Can a classifier sort medical research findings the way expert reviewers do, if no single study appears in both training and testing?
- How early can we detect that a satellite encounter will become dangerous, using only data available at that time?
- Which customers respond to a marketing offer, and does targeting them beat sending it to everyone?
- Can one traffic camera image predict incident likelihood, waiting time and emissions?
- How much does a detection method miss when the telescope samples less often?
Each starts with a domain question, not a model.
Where to find free datasets
- UCI Machine Learning Repository: a long-running collection of research datasets with documentation.
- Kaggle Datasets: a large community collection. Check the source and license of each dataset.
- Data.gov: US government open data across health, transport, climate and more.
- World Bank Open Data: development indicators for nearly every country.
- Domain archives: PhysioNet for physiological signals, NASA and MAST for astronomy, NCBI for genomics.
A good dataset has a clear source, documentation of how it was collected, and enough examples of the cases you care about.
The rigor checks reviewers look for
These are the mistakes that sink student data science papers, and the fixes that strong papers use.
1. Data leakage
Leakage happens when information from the test set sneaks into training, so the model looks better than it is. Common causes: the same patient or study in both sets, features computed using future data, or preprocessing on the full dataset before splitting. Fix it by splitting by group (patient, study, event) and by time when the task is forecasting.
2. Misleading metrics on imbalanced data
When the important class is rare, accuracy and even AUC can look excellent while the model misses every case that matters. Report recall on the rare class, precision, and a confusion matrix.
3. No baseline
A result means little without a comparison: a simple model, a standard rule, or the best published result.
4. One lucky split
Use cross-validation, repeated where possible, and report the spread, not just the best run.
5. Overconfident predictions
In medicine and safety, a model should know when it might be wrong. Calibration and uncertainty estimates turn a classifier into something a person could trust.
6. Results without explanation
Feature importance, error analysis and examples of failures show reviewers you understand your model.
Data science papers by YRI Fellows
These Fellows wrote first-author data science papers with PhD mentors. Many of them built their contribution around exactly the rigor checks above. More are on our published work page.
- Aditya S., an 11th grader, built a two-step neural network for anomaly detection on IBM Cloud telemetry covering 39,365 data points and more than 117,000 features. It detected 16 of 25 anomalies against 6 of 25 in prior research, and he won Best Oral Presentation at IEEE AAIML 2026 in Tokyo. Read his story or the Tokyo announcement.
- Harleen B. showed that a conventional endometriosis model scored an AUC of 0.959 but caught none of the high-risk patients, then fixed it with oversampling and a confidence-weighted ensemble that identified 4 of 5 high-risk patients in held-out data. Read her story.
- Athena R. classified 1,422 medical literature records on GLP-1 drugs and binge eating, using study-grouped cross-validation so no study leaked between training and testing. Her model reached an F1 score of 0.894 on appetite and satiety outcomes. Read her story.
- Krishik S., a 9th grader, used only data available two days before closest approach across 10,460 satellite conjunction events, and showed 25% of eventually high-risk cases hide below standard alert thresholds. Read his story.
- Akhilesh K. benchmarked immunotherapy response models with leakage-free repeated cross-validation and tested whether they could flag their own unreliable predictions. Read his story.
- Saanvi K. built uplift models that estimate each individual's response to an intervention, finding $6.50 more incremental spending per customer with personalized targeting. Read her story.
- Aarav B., a 9th grader, combined YOLO object detection with a Random Forest to predict traffic incident likelihood, waiting time and emissions from a single image. Published in IEEE Xplore. Read his story.
- Sai P., an 11th grader, cataloged spike protein mutations across Omicron lineages and identified 132 immune escape mutations. Accepted at IEEE RCSM 2025. Read her story.
Where student data science papers get published
IEEE conferences. IEEE runs many conferences on machine learning, data analytics and applied computing. Accepted papers appear in IEEE Xplore, as with Aditya's and Aarav's papers.
Domain conferences and journals. A health data paper can go to a medical computing venue, an astronomy analysis to an astronomy venue. Matching the venue to the question often helps acceptance.
Springer and Elsevier. Both publish data science proceedings and journals with peer review.
Code and data sharing. Many reviewers expect to see code. A clean public repository makes your work easier to trust and cite.
Data science at science competitions
Regeneron ISEF has no data science category. Enter the category that matches your question:
- Robotics and Intelligent Machines has a Machine Learning subcategory.
- Software Design covers algorithms and databases.
- Mathematics has a Probability and Statistics subcategory.
- Computational Biology and Bioinformatics fits genomics and computational epidemiology.
- Translational Medical Science fits disease detection and diagnosis.
Data competitions on platforms such as Kaggle build skill, but a leaderboard rank is not a research paper. For project starters, see our machine learning project ideas.
Frequently asked questions
Do I need to know how to code for data science research?
Yes, at a basic level. Python with standard libraries covers most projects, and free notebooks such as Google Colab run in a browser. Several Fellows started with little or no coding and learned during the project.
How is data science research different from a Kaggle project?
A Kaggle project optimizes a score on a fixed dataset. Research asks a question, tests whether the result holds up, compares against prior work, and explains why.
What math do I need?
Algebra and introductory statistics to start. Probability, linear algebra and calculus help as projects get more advanced.
What is the best first data science project?
Pick a field you care about, find one well-documented dataset, and ask whether an existing method fails on an important group. Testing for failure often produces a cleaner paper than chasing accuracy.
Can data science research help with college applications?
Yes. A first-author paper shows you can work with data, think critically about evidence and write clearly, which matters for any major, not only computer science.
How the YRI Fellowship helps you publish data science research
The YRI Fellowship matches you 1:1 with a PhD mentor for 10 weeks of weekly sessions, about 5 to 8 hours of work a week. Your mentor helps you choose a question and dataset, design an evaluation that survives review, and interpret the results. An editing team works through every draft, and a publishing team gets the paper into IEEE, Springer or Elsevier journals and conferences, with free help until it is published.
87% of Fellows publish and 73% place at science competitions. ISEF and JSHS preparation and a mentor letter of recommendation are included. No research experience is needed, and decisions come within 48 hours.
Tuition is $2,997, or 3 payments of $999, with scholarships considered case by case. See pricing and more student stories.
Sources
- Society for Science: ISEF categories and subcategories
- UCI Machine Learning Repository
- Kaggle Datasets
- Data.gov
- World Bank Open Data
- PhysioNet
- Google Colab FAQ
- IEEE Xplore: Anomaly detection in cloud systems, Aditya S.
- IEEE Xplore: Urban traffic incident prediction, Aarav B.
- YRI case study: Harleen B.
- YRI case study: Athena R.
- YRI case study: Krishik S.
- YRI case study: Akhilesh K.
- YRI case study: Saanvi K.
- YRI case study: Sai P.
- YRI Published Work
