Unsupervised Learning in Python

Course Overview

Most machine learning courses start with labeled data and a target to predict. This one starts with the opposite situation. You have a pile of data with no answers attached, and you want to find structure in it. Examples include grouping customers by behavior, sorting articles by topic, and spotting which stocks move together.

The course takes you through four core techniques: clustering, hierarchical clustering with t-SNE visualization, PCA, and non-negative matrix factorization (NMF). You work with real datasets, including grain measurements, fish measurements, company stock prices, Wikipedia articles, and listening data for musical artists. The final chapter combines what you've learned into a small recommender system. It suggests similar articles, and then musical artists, based on the patterns the model finds.

What You Will Learn

• Run k-means on unlabeled data and judge the result with inertia plots and cross-tabulations against known categories.
• Scale and standardize features so that clusters aren't dominated by one large-valued column.
• Build dendrograms with hierarchical clustering, compare linkage methods, and extract cluster labels at a chosen height.
• Use t-SNE to flatten high-dimensional data into a 2D map you can inspect by eye.
• Decorrelate data with PCA and use explained variance to estimate how many dimensions the data really needs.
• Cluster text by applying a PCA variant to tf-idf word-frequency arrays.
• Use NMF to break documents into topics and images into recurring parts.
• Apply cosine similarity to NMF features to produce "more like this" recommendations.

Course Structure

1. Clustering for Dataset Exploration: k-means, choosing the number of clusters, evaluating results, feature scaling, and clustering stocks by price movement.
2. Visualization with Hierarchical Clustering and t-SNE: dendrograms, linkage choices, pulling out intermediate clusterings, and a t-SNE map of the stock market.
3. Decorrelating Your Data and Dimension Reduction: PCA, intrinsic dimension, and clustering Wikipedia articles from tf-idf data.
4. Discovering Interpretable Features: NMF on articles and on images, a comparison with PCA, and the recommender-system exercises.

Who Is This Course For?

It suits people who have already trained a few supervised models in scikit-learn and want to learn the other half of the toolkit. Analysts who need to segment data without labels will also find it useful, as will anyone exploring a new dataset who wants to see its structure before modeling.

It's a poor fit if you want a theory-heavy treatment of the underlying math. It also won't help if you're looking for deep learning approaches such as autoencoders, because the course stays with classical scikit-learn and SciPy methods.

Format & Time Commitment

The course runs about four hours. It combines short videos, which have a transcript you can open, with coding exercises. Each chapter starts with a concept video and then moves into practice problems, and the exercises are worth 4,150 XP in total. A glossary and the datasets are available as resources alongside the lessons. The course was last updated in December 2025.

Finishing it also earns a statement of accomplishment. If you want CPE credits (2.8 are on offer), you need to complete the course and score at least 70% on the qualifying assessment.

Pros and Cons

Pros

• The topics build on each other. Clustering leads into hierarchy and visualization, then dimension reduction, then interpretable features.
• The datasets are realistic: stock prices, text, and music listening data, not just toy examples.
• It ends with a recommender project that ties the techniques together.
• The code is all scikit-learn and SciPy, so it carries over to everyday work.

Cons

• Four hours for four large topics means breadth over depth. You'll learn how to use each method, but not how to tune it in hard cases.
• The exercises are guided, so you're mostly filling in prepared code. You'll still need to practice on your own data to make the skills stick.
• It's a prerequisite-driven course: the page lists a supervised learning course first, which can slow down people who are new to scikit-learn.
• The credential is a statement of accomplishment, not a proctored certification, so it carries less weight with employers than a formal exam would.

FAQ

Do I need to know machine learning first? The page lists Supervised Learning with scikit-learn as a prerequisite. Its FAQ also says comfortable basic-to-intermediate Python is the main requirement and no prior unsupervised learning is needed. Knowing scikit-learn's fit/transform pattern will make the course smoother.

Do I get a certificate? Yes, a statement of accomplishment is included. CPE credits need a separate 70% score on the qualifying assessment.

How long does it really take? The listed time is 4 hours. Plan for more if you want to redo exercises or try the techniques on your own data.

What's the difference between clustering and dimension reduction here? Clustering sorts samples into groups. Dimension reduction (PCA, NMF) re-expresses the data along fewer or more meaningful axes. The course teaches both and shows how they work together.

If this matches the skills you're missing, the course page on DataCamp lets you look at the full exercise list and begin the first chapter.