Skip to content
Home/Resources/AI & ML/Data science projects for beginners
Concept ยท AI & ML

25 Data Science Projects for Beginners, With a Free Dataset, a Question and a Stand-Out Upgrade for Each

By Younus FardeenUpdated 11 Nov 202610 min read
Short answer

Start with a clean dataset and one clear question, then move to messier data and a result someone can open. Below are 25 data science projects in three levels, each with a free dataset whose licence was read on 11 November 2026, the question it answers, tools and one upgrade. Finish three to five properly rather than 25 quickly.

This list is for students and career switchers who need a portfolio. Each project names a dataset, a link, its licence and one question to answer. Every dataset page was opened on 11 November 2026 and its licence or terms read, and sizes come from the dataset pages or their APIs.

One rule runs through the list: a project is a question, not a model. Pick a question, show how you checked the answer and say what could be wrong. Datasets can change or move, so re-check a licence before you publish.

How to pick and finish a project

  1. Pick one question you can say in a sentence.
  2. Load the data and look at it before modelling: missing values, ranges, duplicates.
  3. Build a simple baseline first, such as a group average or the most common class, then beat it.
  4. Split by time or by group where the data needs it, not always at random.
  5. Write what you found, what could be wrong and what you would do next. Put it on GitHub with a short README.

The block below needs only the Python standard library. It loads a 10-row table written inline (made-up numbers, to show the pattern, not a real study) and answers a question: do students who practise more score higher? It also splits by city to check that one city is not driving the result.

import csv
import io
from statistics import mean

# A tiny made-up table: weekly practice hours and test marks for 10 students.
DATA = """student,hours,marks,city
S1,2,48,Pune
S2,3,52,Delhi
S3,4,61,Pune
S4,5,58,Jaipur
S5,6,70,Delhi
S6,7,74,Pune
S7,8,79,Jaipur
S8,9,83,Delhi
S9,10,88,Pune
S10,12,91,Jaipur
"""

rows = list(csv.DictReader(io.StringIO(DATA)))
hours = [float(r["hours"]) for r in rows]
marks = [float(r["marks"]) for r in rows]

def pearson(x, y):
    mx, my = mean(x), mean(y)
    num = sum((a - mx) * (b - my) for a, b in zip(x, y))
    den = (sum((a - mx) ** 2 for a in x) * sum((b - my) ** 2 for b in y)) ** 0.5
    return num / den

# Question: do students who practise more score higher?
low = [m for h, m in zip(hours, marks) if h < 6]
high = [m for h, m in zip(hours, marks) if h >= 6]
print(f"rows loaded: {len(rows)}")
print(f"mean marks, under 6 hours a week: {mean(low):.1f}")
print(f"mean marks, 6 hours or more:      {mean(high):.1f}")
print(f"correlation of hours and marks:   {pearson(hours, marks):.2f}")

# Same question, split by city, to check one city is not driving the result
for city in sorted({r["city"] for r in rows}):
    sub = [r for r in rows if r["city"] == city]
    r_city = pearson([float(r["hours"]) for r in sub], [float(r["marks"]) for r in sub])
    print(f"{city}: {len(sub)} students, correlation {r_city:.2f}")

Output: rows loaded: 10, then mean marks of 54.8 under 6 hours a week and 80.8 for 6 hours or more, a correlation of 0.98, and per-city correlations of 1.00 for Delhi, 0.97 for Jaipur and 1.00 for Pune. A high correlation does not prove practice causes the marks, and 10 rows is far too few for a real conclusion. Real projects need more data and a check on what else differs between students. If pandas is installed the same steps take fewer lines, and the library choices are covered in Python data project ideas.

Beginner projects, 1 to 8

Project and questionDataset and licenceTools and stand-out upgrade
1. Iris species classifier. Can four flower measurements separate three species?UCI Iris, 150 rows, CC BY 4.0pandas, scikit-learn. Compare three models with cross-validation. A tutorial staple, so use it as a warm-up, not a portfolio piece.
2. Wine quality. Which lab measurements best predict a quality score?UCI Wine Quality, 4,898 samples, CC BY 4.0pandas, seaborn. Run it as regression and as classification, and say which is more honest.
3. Student performance. Which habits and background factors go with final grades?UCI Student Performance, 649 rows, 30 features, CC BY 4.0pandas. Drop earlier grades and compare the fit, then write what the data cannot say about causes.
4. SMS spam filter. Can word patterns flag spam messages?UCI SMS Spam Collection, 5,574 messages, CC BY 4.0scikit-learn, TF-IDF, naive Bayes. Study the false positives, the real messages wrongly blocked.
5. India pincode explorer. How are post offices spread across states and districts?data.gov.in All India Pincode Directory, Government Open Data License - IndiaSQL, pandas. Clean the file, load it into SQLite, write ten queries and map the result.
6. India against peers. How has GDP per capita moved for India and similar economies?World Bank Open Data, CC BY 4.0. Its API returned India's 2025 GDP per capita as about US$2,702 on 11 Nov 2026requests, matplotlib. Add a second indicator, such as life expectancy, and test whether the link holds.
7. Is your city getting hotter? How did daily maximum temperature change over five years?Open-Meteo Historical Weather API, free for non-commercial use under CC BY 4.0, under 10,000 calls a dayrequests, pandas. Compare two cities and plot monthly differences from the average.
8. Bike sharing demand. How do hour, season and weather change rentals?UCI Bike Sharing, 17,389 hourly rows, CC BY 4.0pandas, scikit-learn. Split by time, forecast the next week and beat a naive baseline.

Intermediate projects, 9 to 18

Project and questionDataset and licenceTools and stand-out upgrade
9. Heart disease risk. Which clinical measurements go with heart disease?UCI Heart Disease, 303 rows, missing values, CC BY 4.0scikit-learn. Handle missing values openly, report recall and calibration, and state it is not medical advice.
10. Income model fairness audit. Does the error rate differ by sex or race?UCI Adult, 48,842 rows, CC BY 4.0scikit-learn. Compute error rates by group, try one fix and document the trade-off.
11. Customer segments. Which customers buy recently, often and in large amounts?UCI Online Retail II, 1,067,371 rows, CC BY 4.0pandas, k-means. Build RFM scores, add a retention heatmap and write three marketing actions.
12. Card default prediction. Which customers will miss next month's payment?UCI Default of Credit Card Clients, 30,000 rows, CC BY 4.0scikit-learn. Handle class imbalance and pick a threshold from the cost of a missed default.
13. Bank marketing. Who is likely to take a term deposit after a call?UCI Bank Marketing, 45,211 rows, CC BY 4.0scikit-learn. Draw a lift chart showing how many calls could be skipped.
14. Shopper intent. Which website sessions end in a purchase?UCI Online Shoppers Purchasing Intention, 12,330 sessions, CC BY 4.0SQL, scikit-learn. Do a funnel analysis in SQL before any model. About 15.5% of sessions end in a purchase.
15. Predictive maintenance. Can sensor readings flag a failure early?UCI AI4I 2020, 10,000 rows, synthetic, CC BY 4.0scikit-learn. Explain each failure type, and say clearly that the data is synthetic.
16. Finance headline sentiment. Is a headline bullish, bearish or neutral?Hugging Face twitter-financial-news-sentiment, 9,543 train and 2,388 validation rows, MITscikit-learn, transformers. Compare a TF-IDF baseline with a small transformer and label 50 Indian headlines yourself.
17. Packaged food nutrition. How do products compare on sugar and salt per 100 g?Open Food Facts, ODbL database, DbCL contents, CC BY-SA imagespandas. Filter to products sold in India, credit the source and respect the share-alike terms.
18. Movie recommender. Which films should a user see next?MovieLens latest small, 100,836 ratings, 9,742 movies, 610 users. Licence: research purposes, credit required, no commercial usepandas, scikit-learn. Compare a popularity baseline with item similarity. Link to the source instead of re-hosting the file, and keep the project non-commercial.

Portfolio-ready projects, 19 to 25

Project and questionDataset and licenceTools and stand-out upgrade
19. Developer survey explorer. How do languages, experience and AI tool use relate?Stack Overflow Developer Survey 2026, 30,903 responses, ODbL 1.0, DbCL contentspandas, Streamlit. Filter to India, check who answered, and treat pay as self-reported.
20. CO2 emissions dashboard. Which countries cut emissions per person, and how fast?Our World in Data CO2 data, CC BYpandas, Plotly. Publish a dashboard and check the definitions of each column.
21. Hindi named-entity recognition. Can a model tag people, places and organisations in Hindi?Hugging Face naamapadam, Hindi: 985,787 train, 867 test, CC0transformers. Fine-tune a small multilingual model on a subset and report F1 per entity type.
22. Amenity access map. Which areas of a city are far from a hospital or EV charger?OpenStreetMap through Geofabrik, India extract 1.6 GB, ODbLgeopandas. Compute distance to the nearest amenity, map it and note gaps in the data.
23. Build your own air-quality dataset. How does Delhi's air quality move by hour and weather?data.gov.in real-time Air Quality Index, GODL-India, plus Open-MeteoPython, a scheduled script. Save a snapshot every hour for weeks and then forecast. Collecting the data yourself is the stand-out part.
24. Fashion-MNIST image classifier. Where does a small network confuse clothing classes?Hugging Face fashion_mnist, 60,000 train and 10,000 test images, MITPyTorch or Keras. Show a confusion matrix and deploy a demo, see hosting an ML demo on Hugging Face Spaces.
25. Household power forecast. How well can next-day electricity use be predicted?UCI Household Electric Power Consumption, 2,075,259 readings, CC BY 4.0pandas, scikit-learn. Resample to daily, backtest on rolling windows and compare with a seasonal baseline.

What to show recruiters

  • Evidence of applied skill. foundit's July 2026 report describes a market that is 'more discriminating about evidence of applied skill', and TeamLease EdTech names portfolios, internships and assessments as gaining weight. A public repository with a clear README is that evidence.
  • One question per project. Start the README with the question and the answer in two lines, then the method.
  • A baseline and an honest limit. Show the simple model you beat and one thing that could make your result wrong.
  • Something a person can open. A dashboard or a Hugging Face Space beats a notebook. Masai's longer list is 25 data science projects for a job-ready portfolio.
  • Licence hygiene. Credit CC BY sources. Keep share-alike obligations in mind for ODbL data. Do not commit files whose terms forbid redistribution, and do not publish personal data.

For more starting points see AI and ML projects for beginners, data analytics projects with AI and the free datasets for projects list. Role guides: the data scientist roadmap.

Frequently asked questions

What are good data science projects for beginners?

Start with small tabular data and one question: wine quality, student performance, SMS spam, bike sharing or the India pincode directory. Each has a free dataset and a clear task. Once you can load, clean, model and explain one, move to customer segments, fairness audits or a forecast.

Where can I find free datasets for data science projects?

Good sources include the UCI Machine Learning Repository, World Bank Open Data, data.gov.in, Open-Meteo, Our World in Data, OpenStreetMap, the Stack Overflow Developer Survey and Hugging Face. Each states its licence, so read it before you publish. See the free datasets list for more.

How many data science projects do I need for a portfolio?

Three to five good ones beat 25 quick ones. Aim for one clean analysis, one prediction with a baseline and honest limits, and one that people can open, such as a dashboard or a demo. Put each on GitHub with a short README that states the question first.

Can I use a dataset for a public portfolio project?

Usually yes, if you follow the licence. CC BY needs credit. ODbL data, such as OpenStreetMap, needs credit and share-alike for derived databases. Some sets, such as MovieLens, allow research use and bar commercial use, so link to the source instead of re-hosting the file.

Do I need pandas to start?

No. The example above uses only the Python standard library to load a CSV and answer a question. pandas makes this shorter and is worth learning soon, but understanding what the numbers mean matters more than the library.

What should a data science project include?

A clear question, a data description with its source and licence, a baseline, a result with a plain-language answer, the limits of the analysis and code someone else can run. Add one thing people can open, such as a dashboard or a Hugging Face Space.

Which data science project is best for a fresher resume?

One that mirrors the job: a customer segmentation or churn-style project for analyst roles, a time-series forecast for planning roles, or an Indian-language model for AI roles. Pick a dataset with real mess, explain your choices and show a result a non-technical reader can follow.

Sources checked

Continue learning

Roadmap

Data scientist roadmap

Maths, Python, scikit-learn, experiments and one deployed model

Open →
Concept

AI vs data science vs ML

How AI, ML and data science differ, with real B.Tech syllabi

Open →
Roadmap

AI engineer roadmap

LLM APIs, RAG, agents and evals: the 2026 role of building on models

Open →