简体中文
EnglishWebsite review
This review may include AI-generated content and is not individually verified. Domain, certificate and other facts come from the observed data on this page.
What is the UCI Machine Learning Repository?
The UCI Machine Learning Repository is a long-running public archive of datasets for machine learning research and teaching, hosted at the University of California, Irvine. Its front page describes it as a service to the machine learning community, currently maintaining 689 datasets, with options to browse or donate a dataset. It is best understood as a catalog: you come to find a dataset, inspect its documentation, and download it for your own experiments.
UCI Machine Learning Repository
What you'll find there
The collection spans classic small benchmarks and larger, messier real-world data. On the page, popular datasets include:
- Iris — 150 instances, 4 features, classification; described as a small classic dataset from Fisher, 1936.
- Heart Disease — 303 instances, 13 features, classification; drawn from four databases (Cleveland, Hungary, Switzerland, VA Long Beach).
- Wine Quality — about 4.9K instances, 12 features, classification and regression; red and white vinho verde samples from northern Portugal.
- Bank Marketing — about 45.21K instances, 17 features, classification; phone-call direct marketing data from a Portuguese bank, predicting term-deposit subscription.
- Breast Cancer Wisconsin (Diagnostic) — 569 instances, 30 features, classification.
Newer additions shown include a large review-count time-series dataset (about 3.06M instances, 2 features, regression) derived from Google Local Reviews and Amazon Reviews, and HLS-CMDS, a heart and lung sound dataset recorded from a clinical manikin with a digital stethoscope.
Who it's for and how it's used
- Students and instructors use small sets like Iris or Breast Cancer Wisconsin to practice classification workflows end to end.
- Researchers and practitioners use larger sets like Bank Marketing or Wine Quality to compare models, test feature engineering, or benchmark regression and classification methods.
- Anyone needing a citable, publicly documented dataset can search by task type, size, and feature count.
Trade-offs to keep in mind
Convenience and history are the main draws: many datasets are well known, widely cited, and easy to load. The trade-off is that popularity can lead to overuse — Iris in particular is so familiar that strong results on it say little about real-world performance. Dataset quality, licensing, and documentation vary by contributor, so check the individual dataset page before relying on it. Some sets are also old or collected for purposes different from yours, which affects how much you should generalize from them.
A practical next step
If you are choosing a dataset, start from your task: classification, regression, or something else. Then filter by size — small sets for quick iteration, larger ones like Bank Marketing for realistic evaluation — and read the dataset's own description and citation before downloading. For a first project, Iris or Breast Cancer Wisconsin is a reasonable starting point; for a more demanding benchmark, look at Wine Quality or Bank Marketing.
How do I download a specific dataset like Iris or Heart Disease from the UCI Machine Learning Repository?
Downloading a specific dataset such as Iris or Heart Disease from the UCI Machine Learning Repository is straightforward: find the dataset page, then use the download link or data folder provided there. The repository hosts hundreds of datasets as a service to the machine learning community, and each dataset has its own page with files and documentation.
Typical steps
- Go to UCI Machine Learning Repository.
- Search or browse for the dataset name, for example "Iris" or "Heart Disease".
- Open the dataset page.
- Look for a Download button, a Data Folder link, or individual file links (often
.data,.csv,.zip, or.names). - Download the data file and the accompanying documentation, usually a
.namesor.txtfile describing attributes, missing values, and citation details.
What to expect for these two examples
| Dataset | Typical contents | Notes |
|---|---|---|
| Iris | 150 instances, 4 features, classification | A small classic dataset from Fisher, 1936, often used for evaluating classification methods. |
| Heart Disease | 303 instances, 13 features, classification | Combines four databases: Cleveland, Hungary, Switzerland, and the VA Long Beach. |
Practical next step
After downloading, open the dataset page's description file first. It tells you which column is the target, how missing values are coded, and how to cite the source. For Iris, the target is the species; for Heart Disease, the target is the presence or absence of heart disease, often coded as a numeric value. If you prefer working in code, many repositories also offer direct file URLs that you can load with a library such as pandas, but check the dataset page for the exact link rather than guessing.
Decision criterion
If you need a quick, clean dataset for testing a classifier, Iris is the easier starting point. If you want a more realistic medical classification task with more features and some missing values, Heart Disease is the better choice.
What should I consider when choosing between popular datasets such as Iris, Wine Quality, or Bank Marketing for a classification project?
Start with the decision you are actually making: these three datasets differ mainly in size, feature type, class structure, and how much preprocessing they demand. They are not interchangeable benchmarks.
Quick comparison
| Dataset | Instances / Features | Task fit | Main trade-off |
|---|---|---|---|
| Iris | 150 / 4 | Multiclass classification (3 species) | Tiny and clean, so results are easy to over-interpret; good for learning and sanity checks, weak for realistic performance claims |
| Wine Quality | 4.9K / 12 | Classification or regression | Real physicochemical measurements, but quality scores are imbalanced and the red/white split matters |
| Bank Marketing | 45.21K / 17 | Binary classification | Realistic scale and mixed feature types, but class imbalance and preprocessing effort are much higher |
How to choose
Iris suits a first pass: verifying that a pipeline, plotting code, or cross-validation setup works. Its four features are numeric and the classes are well separated, so almost any reasonable method scores well. That is also its limitation — a high accuracy here says little about harder problems. Use it to debug, not to compare models you care about.
Wine Quality sits in the middle. The UCI entry describes two datasets of red and white vinho verde samples, with quality modelled from physicochemical tests. If you want to practise feature scaling, handling ordinal or imbalanced target values, and deciding whether to pool or separate the two wine types, this is a useful step up. The class distribution is uneven, so accuracy alone can mislead; look at per-class recall or a confusion matrix.
Bank Marketing is the closest to a real tabular project. The UCI description covers direct marketing phone campaigns from a Portuguese bank, with the goal of predicting whether a client subscribes to a term deposit. Expect mixed numeric and categorical columns, many more rows, and a minority positive class. This is the right choice if your goal is to demonstrate realistic handling of imbalance, encoding, and threshold selection — but budget time for cleaning and validation.
A practical next step
Pick based on the skill you want to test, not the dataset's fame. If you are learning the workflow, start with Iris and move to Wine Quality. If you are building a portfolio piece or evaluating a model for deployment-style use, choose Bank Marketing and report precision, recall, and a baseline alongside accuracy.
You can browse and download all three from the UCI Machine Learning Repository, which lists instance and feature counts for each entry so you can confirm the current versions before you start.
How can I donate or contribute a new dataset to the UCI Machine Learning Repository?
You can contribute a dataset through the repository's "Contribute a Dataset" link, which appears on its homepage alongside the option to browse existing datasets. The repository currently maintains 689 datasets as a service to the machine learning community, so new contributions are part of an established, ongoing collection rather than a one-off upload.
UCI Machine Learning Repository
H3 What a contribution involves
- A dataset file or archive, plus a clear description of what it contains.
- Documentation of the features, instance counts, and the task type (for example classification or regression).
- Enough context for someone else to understand how the data was collected and what the variables mean.
H3 How to judge whether your dataset fits Look at the popular datasets listed on the site as a rough guide. Iris is a small classic dataset with 150 instances and 4 features, used for evaluating classification methods. Heart Disease combines four databases and has 303 instances and 13 features. Wine Quality covers red and white vinho verde samples with 4.9K instances and 12 features, for classification and regression. Bank Marketing has 45.21K instances and 17 features for a classification goal. Breast Cancer Wisconsin (Diagnostic) has 569 instances and 30 features. These range from tiny to fairly large, so size alone is not the barrier; clear documentation and a well-defined task matter more.
If your data is messy, undocumented, or restricted by privacy or licensing terms you cannot share, it is unlikely to be accepted in its current form. If it is clean, ethically shareable, and comes with a description of features and collection method, it is a reasonable candidate.
H3 A practical next step Before submitting, write a short README covering the source, the meaning of each column, the number of instances and features, and the intended task. Compare it against the documentation style of an existing entry such as Iris or Breast Cancer Wisconsin (Diagnostic) to see whether a stranger could use your data without asking you questions. That README is usually the difference between a dataset that gets used and one that sits unread.
What are the licensing and citation requirements for using datasets from the UCI Machine Learning Repository in my research?
Licensing at UCI is dataset-specific, not repository-wide. The page describes the repository as a service to the machine learning community where you can "donate and find datasets," but it does not state a single blanket license covering all 689 datasets. In practice, each dataset page carries its own terms — often a Creative Commons license, sometimes a citation requirement, occasionally restrictions from the original collector or publisher.
What this means for your research
- Check the individual dataset page first. The license and any citation request appear on that dataset's page, not on the repository home page. Two datasets from the same repository can have different terms.
- Cite the dataset, not just the repository. Most dataset pages list a preferred citation, usually the original paper or the donor's suggested reference. Use that exact reference where provided.
- Attribute the source. Even when a license is permissive, standard practice is to credit the UCI Machine Learning Repository as the distribution source and name the dataset and version.
- Watch for non-commercial or no-redistribution clauses. Some donated datasets inherit restrictions from the original study or data provider. These can limit commercial use or sharing of the raw files.
Citation format that usually works
[Authors/Donors]. ([Year]). [Dataset Name] [Dataset]. UCI Machine Learning Repository. UCI Machine Learning Repository
Replace the bracketed parts with what the dataset page specifies, and add the DOI or dataset URL if one is listed.
Practical decision criterion
If your intended use is commercial, or if you plan to redistribute the data, treat the dataset page's license text as controlling and confirm it permits that use before you build on it. For academic publication with no redistribution, a clear citation plus a link to the dataset page is normally sufficient.
Next step
Open the specific dataset's page, look for a "License," "Citation," or "Source" section, and copy the required citation into your references as you download the files — retrofitting citations later is the most common source of errors. For datasets that originated elsewhere, such as Wine Quality or the reviews datasets, also check the original source the page points to, since its terms may be stricter.
How does the UCI Machine Learning Repository compare to other dataset sources like Kaggle for finding machine learning data?
UCI Machine Learning Repository is a long-standing academic archive of curated benchmark datasets, while Kaggle Datasets is a community platform built around competitions, notebooks and a much broader, more uneven catalogue. They overlap in purpose but suit different jobs.
UCI Machine Learning Repository works best when you want a small, well-documented dataset to test an algorithm, reproduce a classic result, or teach a method. Its popular entries are exactly that: Iris (150 instances, 4 features, classification), Heart Disease (303 instances, 13 features), Wine Quality (about 4.9K instances, 12 features, classification and regression), Bank Marketing (about 45.2K instances, 17 features) and Breast Cancer Wisconsin (Diagnostic) (569 instances, 30 features). Most are tabular, moderate in size, and come with citations and descriptions you can reference directly. The trade-off is scale and freshness: it maintains hundreds of datasets rather than tens of thousands, so it is not the place to look for large-scale image, text or production-like data.
Kaggle's strength is variety and tooling. You get many user-uploaded datasets, often large or domain-specific, plus in-browser notebooks, discussion and competition infrastructure. The cost is variable provenance: licences, cleaning, documentation and duplicates need checking before you rely on a dataset.
H3 Quick comparison
| Criterion | UCI | Kaggle |
|---|---|---|
| Curation | Reviewed, academic, benchmark-oriented | Community-uploaded, variable quality |
| Typical size | Small to medium tabular | Small to very large, many modalities |
| Documentation | Dataset descriptions, citations | Varies by uploader |
| Extras | Stable identifiers for papers and teaching | Notebooks, competitions, discussion |
| Best for | Reproducible experiments, coursework, baselines | Prototypes, competitions, niche or large data |
A practical rule: start with UCI when you need a defensible baseline or a dataset you can cite in a paper or assignment; move to Kaggle when you need volume, a specific domain, or a leaderboard to measure against. If you are choosing a dataset for a portfolio project, pick one where the documentation matches your goal — a small, clean UCI set lets you focus on modelling and evaluation, while a messy Kaggle set lets you show data cleaning and feature engineering.
Proofpoint
User reviews (0)