Machine Learning Project Ideas for High School Students: 40 Research Questions With Datasets

Machine learning is the most accessible research field for a high school student: the data is public, the tools are free, and a laptop is enough. It is also the field where students most often produce something that looks like research and is not. Training a model on a Kaggle dataset and reporting the accuracy is a tutorial, not a project.

The difference is the question. Every idea below is written as a research question with a dataset and a baseline, because a question with a baseline is something you can be wrong about, and being able to be wrong is what makes it research. Ideas marked with an asterisk are ones that, done properly, have the shape of a publishable paper or an ISEF entry.

How to tell a project from a tutorial

Before the list, the test. A machine learning project is research if it has all four:

1. A specific question whose answer is not already in the dataset's documentation.
2. A baseline you are trying to beat or explain: a published method, a simple model, or a standard benchmark number.
3. An honest evaluation with held-out data, confidence intervals or multiple runs, and a comparison the reader can trust.
4. A finding that would be worth a paragraph to someone in the field even if the number is small.

If you cannot name the baseline, you have a tutorial. What Makes a Project ISEF-Worthy applies the same test to any field.

Health and medicine

1. * Can a small model trained on a public ICU dataset (MIMIC-IV demo, PhysioNet) predict 30-day readmission as well as the published logistic regression baselines, and which features drive it?
2. * Does adding simple text features from discharge notes improve mortality prediction over structured data alone, on the MIMIC demo subset?
3. Can wearable accelerometer data from a public dataset classify sleep stages with accuracy comparable to published methods using a model small enough to run on a phone?
4. * Does a model trained to detect diabetic retinopathy on one public fundus dataset hold its accuracy on a second dataset from a different country, and what explains the drop?
5. Can a time-series model on public continuous glucose monitor data predict hypoglycemia 30 minutes ahead better than a persistence baseline?
6. * Which symptoms in a public COVID or flu self-report dataset best predict a positive test, and does a transparent model match a black-box one?

Climate and environment

7. * Can satellite imagery from Sentinel-2 classify urban tree canopy at the block level, validated against city tree inventories?
8. Does a model trained on historical air quality and weather data predict next-day PM2.5 better than the persistence and seasonal baselines for your city?
9. * Can machine learning on NOAA buoy data detect the onset of marine heatwaves earlier than the standard threshold definition?
10. Which features in public wildfire datasets best predict fire spread in the first 24 hours, and how well does a simple model do against the published ones?
11. * Can a model on public river gauge and rainfall data forecast flood stage 12 hours ahead for a specific basin, compared with the agency's baseline?

Language and text

12. * Can a small fine-tuned language model match a large general model on a narrow classification task, such as detecting misinformation in a specific domain, at a fraction of the compute?
13. Does a model trained to detect toxic comments on one platform's public dataset transfer to another platform, and what kinds of text break it?
14. * How does the sentiment of news coverage of a topic, measured by a classifier on a public news dataset, track against a real-world indicator over time?
15. Can readability of science articles be predicted better by a learned model than by standard formulas like Flesch-Kincaid, validated on human ratings?
16. * Do large language models show measurable bias when generating descriptions of people from different groups, quantified with a controlled prompt set and a published bias benchmark?

Vision

17. Can a model trained on a public plant disease dataset identify disease from photos taken on a phone in a real garden, and how far does accuracy fall from the benchmark?
18. * Does data augmentation designed to mimic distribution shift improve a classifier's robustness on a standard shift benchmark, measured honestly against the un-augmented baseline?
19. Can a lightweight model count pollinators in time-lapse images of a flowering plant accurately enough to replace manual counting?
20. * Can a model detect pedestrian crossing hazards in public dashcam datasets, and does performance differ at night or in rain?

Education and social science

21. * Which features in a public student performance dataset predict course failure early enough to intervene, and does a transparent model match a complex one?
22. Does a model on public school district data predict chronic absenteeism, and which modifiable factors matter most?
23. * Can a survey experiment's results be predicted from a pre-registered model, testing whether message framing effects replicate across two public datasets?
24. How accurately can a model predict a song's chart performance from audio features in a public music dataset, and what does the failure pattern say about taste?

Science and engineering

25. * Can a graph neural network predict molecular properties on a standard benchmark as well as published models when trained with a fraction of the data?
26. Does a model trained on protein sequence alone predict thermostability competitively with published methods?
27. * Can transit timing variations in public TESS light curves be detected with a machine learning model, and does it recover known cases?
28. Can a model predict battery degradation from early cycle data in a public battery dataset better than the published baseline?
29. * Which materials features best predict band gap in a public materials database, and does a simple model generalize to a held-out family of compounds?
30. Can a physics-informed neural network solve a simple fluid problem more accurately than a standard network with the same data?

Fairness, interpretability and evaluation

31. * Does a model trained for a public hiring or lending prediction task show disparate error rates across groups, and which mitigation technique reduces the gap at the lowest accuracy cost?
32. How much of a model's reported accuracy on a popular benchmark survives when the test set is de-duplicated against the training set?
33. * Do explanation methods such as SHAP and integrated gradients agree on the same model, and when they disagree, which one matches a ground truth you can construct?
34. How sensitive is a published benchmark result to random seed, and what does that imply about claimed improvements smaller than the variance?

Audio and signals

35. Can a model detect a specific bird species in public field recordings with accuracy comparable to published baselines, and how much labeled data does it actually need?
36. * Can heart sound recordings in a public dataset be classified as normal or abnormal by a small model suitable for a low-cost stethoscope?
37. Does a wake-word detector running on a microcontroller reach usable accuracy within a fixed power budget?

Systems and efficiency

38. * How much can a model be compressed by pruning and quantization before accuracy on a benchmark drops measurably, and does the answer differ by task?
39. Does a modified learning rate schedule change the final accuracy or only the training time, measured across multiple seeds?
40. * Can a small model distilled from a large one match it on a narrow task, and what is the smallest size at which that holds?

Turning an idea into a paper

The students who publish machine learning work do five things that tutorials skip:

• They read the three most relevant papers first and borrow the evaluation protocol, so their numbers are comparable. How to Read a Scientific Paper covers how.
• They fix the baseline before training anything. The baseline is the paper's honesty.
• They run everything with multiple seeds and report the spread. A single run is an anecdote.
• They write the limitations section seriously. Reviewers trust a paper that knows what it did not show.
• They pick a venue that fits. IEEE student and regional conferences, workshop tracks at larger conferences, and a short list of real journals accept strong high school work. Journals That Publish High School Research ranks them.

A mentor who has published in machine learning collapses the time this takes from a year to ten weeks, mostly by knowing which questions are already answered and which evaluation mistakes get a paper rejected. That is the first thing a YRI Fellowship mentor does with a student in this field, and it is why Fellows' machine learning papers have been accepted at IEEE conferences in the same year they started.

Frequently Asked Questions

What are good machine learning project ideas for high school students? Projects framed as a research question with a public dataset and a baseline to beat, for example predicting ICU readmission on the MIMIC demo data, classifying tree canopy from Sentinel-2 imagery, or testing whether a small fine-tuned model matches a large one on a narrow task. The list above has 40 across health, climate, language, vision, education and science.

Can a high school student publish a machine learning paper? Yes. Students publish in IEEE conferences, workshop tracks and legitimate journals every year, usually with a mentor and a project that has a real question, an honest evaluation with multiple runs, and a comparison to a published baseline.

What datasets can high school students use for machine learning? Public sources include PhysioNet and the MIMIC demo for health, NOAA and Sentinel for climate, Kaggle and the UCI repository for general data, Hugging Face for text, and TESS and materials databases for science. The dataset matters less than the question asked of it.

Do you need a GPU for a high school machine learning project? Usually not. Most of the ideas above run on a laptop or on free cloud notebooks. Projects that need more compute can be scoped to a subset of data or a smaller model, which is often a better research question anyway.

How is a machine learning research project different from a Kaggle competition? A competition optimizes one number on one dataset. A research project asks a question, compares against a baseline, evaluates honestly, and reports a finding with its limitations. The same dataset can support either; the framing decides which one you did.