Google Introduction to Clustering: Free Machine Learning Course

Unlabelled data can contain useful patterns without telling you what those patterns mean. Google’s Introduction to Clustering explains how to group similar examples, choose a meaningful measure of similarity and judge whether the resulting groups are useful.

This free, self-paced machine learning course focuses on k-means and the decisions around it. You study the workflow from preparing features to evaluating clusters, then explore how neural-network embeddings can represent similarity in more complex data. It is a conceptual course with visual examples and knowledge checks, rather than a full coding bootcamp.

Course at a glance

What you’ll learn

• Explain how clustering groups unlabelled examples and differs from classification.
• Recognise common clustering approaches and the course’s focus on k-means.
• Prepare numerical features so differences in scale do not distort comparisons.
• Follow the assignment and centroid-update steps of k-means.
• Evaluate cluster sizes, distances and the usefulness of the results.
• Compare manual similarity measures with measures based on learned embeddings.

Skills you’ll gain

• Unsupervised learning
• Feature scaling
• K-means reasoning
• Similarity measures
• Cluster evaluation
• Embedding concepts

Start with the meaning of similarity

The opening lessons connect clustering with applications such as market segmentation and search-result grouping. The central question is what makes two examples similar. That definition is a design decision: the features you choose and how you combine them influence the groups an algorithm will produce.

Data preparation therefore comes before clustering. You review scaling, normalisation, transformations and quantiles, along with the problem of missing values. The visual examples make a practical point: changing the scale of a feature can change which examples appear close to one another.

Follow k-means without losing sight of its limits

The k-means section walks through selecting a number of clusters, choosing initial centroids, assigning examples and updating those centroids. Following that loop helps you understand what the algorithm is optimising and why different starting positions can affect its results.

The course also explains its weaknesses. Unequal densities, outliers and high-dimensional data can make straightforward k-means a poor fit. Learning those limits matters as much as learning the steps: a neat-looking grouping is not enough to establish that the method suits the data.

Evaluation looks at cluster cardinality, cluster magnitude and downstream performance. You also consider how increasing the number of clusters changes the total distance to centroids. These are ways to investigate quality and choose a useful level of grouping, rather than automatic proof that one answer is correct.

Compare hand-designed measures with embeddings

For a small dataset with understandable features, a manual similarity measure can be easier to inspect. The course then introduces a different route for more complex features: use a neural network to create lower-dimensional embeddings and compare those representations.

You learn the roles of an autoencoder and a predictor network, followed by Euclidean distance, cosine similarity and dot product. The lessons show how vector length can influence comparisons. This gives you a more precise way to discuss why changing the measure might change the grouping.

Study actively before adding code

Google expects knowledge equivalent to its Problem Framing and Machine Learning Crash Course material, including numerical data and generalisation. Familiarity with vectors and basic ML terminology is helpful for the later sections.

For optional personal practice, sketch a few fictional items with two numerical features. Describe what similarity should mean, trace one k-means update and identify an outlier that might pull a centroid away from the main group. This is an independent study exercise, not a Google assignment or a claim that you have trained a model.

Free access and realistic expectations

The advertised learning route is public reading and browser knowledge checks. It does not require Colab, a cloud subscription, a payment card or paid software. Follow the course’s sequential navigation to reach the current lessons. No certificate, badge or academic credit is promised.

Use the course to develop your understanding of clustering decisions and terminology. Building and validating a system on a real dataset would require additional implementation and testing; the 110-minute estimate does not promise production-ready expertise.

Enroll Now ↗

Frequently asked questions