Machine Learning Projects for Final Year Students: 10 Dataset Ideas

Last updated: 13 October 2026 · By the Asuraa Team

Quick answer: Good final year machine learning projects pair a documented public dataset with a clear metric, a baseline and a simple demo. Ten options: crop recommendation, UCI heart disease (303 Cleveland patients), CarDekho used car prices, credit card fraud, Stanford IMDB sentiment (25,000 train and 25,000 test reviews), ISOT fake news, a MovieLens recommender, household energy forecasting, PlantVillage disease detection and UCI Devanagari characters (92,000 images, 46 classes).

Key takeaways

• Final year ML projects are judged on a clear metric, a fair baseline, honest validation and a working demo, not only on accuracy.
• The UCI Devanagari Handwritten Character Dataset contains 92,000 32x32 grayscale images across 46 classes, split 85% training and 15% testing.
• The Stanford Large Movie Review Dataset provides 25,000 training and 25,000 test reviews with balanced sentiment labels.
• The UCI Heart Disease data is most often used through its Cleveland subset of 303 patients, which suits careful small-data validation.
• MovieLens Latest Small has 100,000 ratings from 600 users on 9,000 movies, enough for a recommender project on a laptop.

Which machine learning projects suit final year students?

The best final year machine learning projects pair a well-documented public dataset with a clear evaluation metric and a simple deployed demo. The ten projects below cover classification, regression, NLP, computer vision, recommendation and forecasting, so a team can pick one that fits its branch and still show end-to-end skills.

If you want to understand the theory before choosing, how to learn machine learning sets out the order of topics.

What does each machine learning project involve, step by step?

Each project lists a goal, steps, what it proves and an example resume bullet. Bracketed parts are placeholders; report only the metrics you actually obtain.

1. Crop recommendation

Goal: recommend a crop from soil and climate inputs. Dataset: Crop Recommendation Dataset, built for recommending suitable crops to maximise yield.

1. Explore feature ranges for each crop.
2. Train logistic regression, random forest and gradient boosting.
3. Compare with stratified cross-validation.
4. Deploy a Streamlit form that returns the top three crops.

Shows: a full classification workflow with an Indian use case. Example resume bullet (example only): "Built a crop recommender with random forest reaching [x]% cross-validated accuracy; deployed as a Streamlit app."

2. Heart disease risk

Goal: predict presence of heart disease and explain the drivers. Dataset: UCI Heart Disease; the Cleveland subset of 303 patients is the one most researchers use, and the target is usually simplified to presence versus absence.

1. Handle missing values and encode categorical fields.
2. Use repeated cross-validation, since the dataset is small.
3. Report recall alongside accuracy.
4. Explain predictions with SHAP values.

Shows: careful validation on small data and interpretability. Example resume bullet (example only): "Trained an interpretable heart disease classifier on 303 patients with [x] recall under repeated cross-validation."

3. Used car price prediction

Goal: estimate a fair resale price for a used car in India. Dataset: Vehicle dataset from CarDekho, used car listings collected from websites.

1. Derive car age and clean brand names.
2. One-hot encode fuel, seller type and transmission.
3. Compare linear regression with gradient boosting on MAE.
4. Build a simple price-estimate form.

Shows: regression and feature engineering. Example resume bullet (example only): "Predicted used car prices with gradient boosting at Rs [x] MAE, [y]% lower than a linear baseline."

4. Credit card fraud detection

Goal: flag fraudulent transactions with few false alarms. Dataset: Credit Card Fraud Detection, anonymised transactions labelled fraudulent or genuine.

1. Measure the class imbalance before modelling.
2. Split by time order to mimic production.
3. Compare class weights with resampling.
4. Choose a threshold from the precision-recall curve and explain the trade-off.

Shows: imbalanced learning and the right metric for the problem. Example resume bullet (example only): "Built a fraud classifier with [x] recall at [y] precision; chose the threshold from a precision-recall analysis."

5. Movie review sentiment

Goal: classify review sentiment and compare classic and transformer models. Dataset: Stanford Large Movie Review Dataset, 25,000 training and 25,000 test reviews with balanced labels; the authors ask users to cite their ACL 2011 paper.

1. Build a TF-IDF and logistic regression baseline.
2. Fine-tune a small pretrained transformer.
3. Compare accuracy, training time and model size.
4. Analyse misclassified reviews such as sarcasm.

Shows: NLP and cost-aware model choice. Example resume bullet (example only): "Compared TF-IDF and fine-tuned transformer models on 50,000 IMDB reviews; transformer gained [x] points accuracy at [y] times inference cost."

6. Fake news detection

Goal: classify news articles as real or fake, and check whether the model learns content or shortcuts. Dataset: ISOT Fake and Real News.

1. Train a baseline text classifier.
2. Inspect top features for source-specific shortcuts such as agency names.
3. Remove those cues and retrain.
4. Report the accuracy drop honestly.

Shows: data-leakage awareness, a sign of maturity. Example resume bullet (example only): "Found [n] leakage cues in a fake news dataset; after removing them, accuracy fell from [x]% to [y]%, giving a realistic estimate."

7. Movie recommender

Goal: recommend films to users from their ratings. Dataset: MovieLens; the Latest Small version has 100,000 ratings from 600 users on 9,000 movies, while larger versions run to 32 million ratings.

1. Start with a popularity baseline.
2. Train matrix factorisation for collaborative filtering.
3. Evaluate with precision@k on held-out ratings.
4. Add a content-based fallback for new users.

Shows: recommender systems and ranking metrics. Example resume bullet (example only): "Built a MovieLens recommender with matrix factorisation improving precision@10 from [x] to [y] over a popularity baseline."

8. Household energy forecasting

Goal: forecast next-day household electricity use. Dataset: UCI Individual Household Electric Power Consumption, 2,075,259 one-minute readings from one house near Paris from December 2006 to November 2010, with about 1.25% of rows missing values.

1. Resample to hourly totals and fill short gaps.
2. Create lag, rolling and calendar features.
3. Compare a seasonal naive baseline, gradient boosting and an LSTM.
4. Backtest with a rolling origin.

Shows: time series discipline. Example resume bullet (example only): "Forecast hourly household power from 2 million readings; gradient boosting cut error [x]% versus a seasonal naive baseline."

9. Plant disease detection

Goal: identify crop diseases from leaf photos. Dataset: PlantVillage, images of diseased plant leaves with labels.

1. Split by image source to avoid near-duplicates across sets.
2. Fine-tune a pretrained CNN such as MobileNet or EfficientNet.
3. Test on a few real field photos, not only the dataset.
4. Export a lightweight model for a mobile or web demo.

Shows: transfer learning and testing beyond lab conditions. Example resume bullet (example only): "Fine-tuned MobileNet on PlantVillage to [x]% test accuracy; documented the drop on [n] real field photos."

10. Devanagari character recognition

Goal: recognise handwritten Devanagari characters. Dataset: UCI Devanagari Handwritten Character Dataset, 92,000 grayscale 32x32 images across 46 classes, split 85% training and 15% testing.

1. Train a small CNN from scratch.
2. Add augmentation such as small rotations.
3. Examine the confusion matrix for similar-looking characters.
4. Build a drawing-pad demo that predicts as you write.

Shows: computer vision with an Indian-language angle. Example resume bullet (example only): "Trained a CNN on 92,000 Devanagari characters across 46 classes to [x]% test accuracy, with a live drawing-pad demo."

How do you turn an ML project into a strong final year submission?

Treat the project like a small research study: state a baseline, compare models fairly and report limitations.

1. Write the problem, metric and baseline before training anything.
2. Fix random seeds and keep a clean train-validation-test split.
3. Log experiments in a simple table: model, settings, metric.
4. Deploy a minimal demo with Streamlit or Gradio.
5. End the report with limitations and ethical concerns, such as bias in health data.

What do most guides on machine learning projects get wrong?

Most guides celebrate high accuracy without asking whether it is meaningful. On imbalanced fraud data, a model that predicts "genuine" for everything scores very high accuracy and catches no fraud. On fake news data, models can learn the source's formatting rather than the content. Examiners and interviewers respect a candidate who finds these problems.

Guides also skip baselines. A transformer's accuracy means little until it is compared with a TF-IDF model that trains in seconds.

Lastly, a model that never leaves a notebook is half a project. A simple demo makes the work easy to show in an interview.

How should you present ML projects to recruiters?

Publish each project on GitHub with a README covering the problem, data, baseline, results table, demo link and limitations. The projects and portfolio guide for India and how to build a portfolio for jobs explain layout; the machine learning engineer resume for freshers shows how projects fit on the page. Prepare for follow-up questions with machine learning interview questions, and check your bullets against a target job with the Asuraa resume reviewer.

Related guides

• Projects and Portfolio Guide for Freshers: Data, Software and AI
• Excel Projects for a Data Analyst Portfolio: 6 Ideas With Datasets
• Web Development Projects for a Fresher Resume: 8 Ideas to Full Stack
• Power BI Projects for a Portfolio: 6 Dashboards With Public Data

FAQ

Which machine learning project is best for a final year student?

Choose one that fits your branch and lets you show the full workflow: problem, baseline, model comparison, validation and a demo. Crop recommendation and plant disease detection suit agriculture interests; Devanagari character recognition suits computer vision; fraud detection suits finance. A well-validated simple model beats an unexplained deep network.

Where can I get datasets for final year ML projects?

The UCI Machine Learning Repository hosts datasets such as Heart Disease, Devanagari Handwritten Characters and Household Power Consumption with clear documentation. Kaggle hosts PlantVillage, CarDekho and fraud datasets. GroupLens publishes MovieLens, and Stanford publishes the Large Movie Review Dataset. Check each licence and citation request.

Is a deep learning project better than classic machine learning?

Not automatically. Deep learning fits image and text tasks such as PlantVillage or Devanagari recognition, but tabular problems like heart disease or car prices are often better served by gradient boosting. Reviewers value a fair comparison against a simple baseline more than the choice of model family.

How do I avoid misleading accuracy in ML projects?

Pick a metric that matches the problem: recall and precision for fraud or disease, MAE for price prediction, precision at k for recommenders. Check for data leakage, such as source cues in fake news text or near-duplicate images across train and test sets, and report results on a held-out test set.

Should a final year ML project be deployed?

Yes, a simple deployment helps a lot. A Streamlit or Gradio app that takes inputs and returns predictions lets examiners and interviewers try your model in seconds. Keep it lightweight, document how to run it and link it from your README and resume next to the project title.

How do I write an ML project on my resume?

State the task, dataset size, model, metric and result, plus any deployment. An example only: Trained a CNN on 92,000 Devanagari characters across 46 classes to [x]% test accuracy with a live drawing-pad demo. Report only metrics you actually obtained and be ready to explain your validation choices.

Final thoughts

Pick one project, set a baseline, validate honestly and ship a small demo; that combination is rarer than high accuracy and far more convincing. If you want a second opinion on scope before you start, a session with an ML practitioner on Asuraa mentors can save weeks.