Data and modeling
Machine learning for unlabeled data
Machine learning can find structure in data with no labels. Clustering groups similar customers or records, dimensionality reduction maps high-dimensional data into a view you can inspect, and autoencoders learn what normal data looks like so unusual records stand out.
How machine learning is used on unlabelled data
Mapping high-dimensional data
Dimensionality reduction turns thousands of measurements per record into a two-dimensional map in which similar records sit together. Scientists and analysts use these maps to spot groups, gradients and odd samples before any labelling.
Real exampleUMAP for single-cell data (Becht et al., Nature Biotechnology, 2019)The authors applied UMAP to mass cytometry and single-cell RNA sequencing datasets and compared it with five other dimensionality-reduction tools.
Finding topics in documents
Topic models describe a collection of documents as mixtures of themes, each a distribution over words, with no labelled examples. They are used to explore archives, research literature and customer feedback.
Real exampleLatent Dirichlet Allocation (Blei, Ng and Jordan, JMLR 2003)The paper introduced LDA, a generative probabilistic model for collections of discrete data such as text corpora.
Learning representations without labels
Self-supervised learning creates its own training task from raw data, such as matching two augmented views of the same image. The learned representation is then fine-tuned with only a small labelled set.
Real exampleSimCLR (Chen et al., 2020)SimCLR is a simple framework for contrastive self-supervised learning of visual representations, without specialised architectures or a memory bank.
Creating labels cheaply
When hand labelling is the bottleneck, teams write rules that label data noisily and use a model to combine and denoise them. The result trains an ordinary classifier.
Real exampleSnorkel (Ratner et al., 2017)Snorkel lets users train models without hand labelling by writing labelling functions, whose noisy outputs it denoises without access to ground truth.
Can you learn machine learning for unlabelled data?
Yes. Unsupervised learning uses the same Python tools as supervised learning, and the core methods, clustering and dimensionality reduction, are taught early. You need some linear algebra for PCA and a clear way to judge results that have no labels.
- Python and data toolsLoad, scale and plot data in NumPy, pandas and Matplotlib, since unsupervised results are mostly judged by inspection.Learn it in Data Tools
- Vectors and matricesLearn distances, dot products and matrix decompositions, which underlie clustering, PCA and embeddings.Learn it in Maths
- Clustering and dimensionality reductionApply k-means, hierarchical clustering, PCA and UMAP, and check whether groups are stable across runs and samples.Learn it in Machine Learning
- Representation learningTrain autoencoders and use self-supervised embeddings, then fine-tune on a small labelled set.Learn it in Deep Learning
Moving into unsupervised and self-supervised learning
If you already work in the field
Researchers and analysts with large unlabelled datasets, such as survey text, sensor logs or biological measurements, already know which groupings would be meaningful. Start by clustering and mapping your own data and checking the groups against what you know.
If you come from data science or ML
If you mostly train supervised models, the shift is in evaluation: with no labels, you judge results by stability, by downstream performance and by domain experts. Learn how the domain describes its own categories before naming clusters.
Roles that do this workData scientistMachine learning engineerResearch scientistComputational biologistNLP engineer
Tools you will use
| Tool | What you use it for |
|---|---|
| scikit-learn clustering | Run k-means, HDBSCAN, Gaussian mixtures and PCA on tabular or embedded data. |
| UMAP | Project high-dimensional data into two dimensions to inspect cluster structure. |
| PyTorch | Train autoencoders and other models that learn representations without labels. |
| pandas | Profile cluster members and compare their feature averages. |
Public data to practise on
- Fashion-MNIST: Greyscale images of clothing in ten categories; hide the labels, cluster, then check the groups against them.
- Wholesale Customers (UCI): Annual spending by product category for a distributor's clients, for customer segmentation.
Your first project
Download Fashion-MNIST, cluster the images with k-means on PCA features while ignoring the labels, and compare the clusters with random assignment using the adjusted Rand index against the true categories.
Common questions
What is unsupervised learning?
Unsupervised learning finds structure in data without target labels, such as groups, low-dimensional maps or unusual points. Clustering and dimensionality reduction are the most common examples.
How do you evaluate a model without labels?
Use internal measures such as silhouette score, check that results are stable across samples, test whether the output improves a downstream task, and ask domain experts to review examples.
What is the difference between unsupervised and self-supervised learning?
Unsupervised learning usually means clustering or density estimation. Self-supervised learning builds a prediction task from the data itself, such as predicting a masked word, and is how large language and vision models are pretrained.
Can I train a classifier with only a few labels?
Often, yes. Semi-supervised learning, weak supervision and fine-tuning a pretrained model can reach useful performance with far fewer labels than training from scratch.
Courses for these skills
- Practical Machine Learning in PythonChoose an anomaly method and interpret its scores.51 published lessons · more lessons in production
- Data Tools: NumPy, pandas, Matplotlib and scikit-learnInspect feature scales and representative rows.23 published lessons · all planned lessons open
- Deep Learning with PyTorchStudy representation-learning objectives and neural training.Syllabus available · lessons open as produced
- Maths for Machine LearningUnderstand latent vectors and distances.26 published lessons · all planned lessons open
- Statistics and Probability for Machine LearningUnderstand tails, robust summaries and base-rate effects.41 published lessons · all planned lessons open
New to programming or maths? Start with Python, or follow the machine learning path in order. Published lessons are free to read; compare Free and Pro.