Unsupervised Learning

Unsupervised learning is machine learning on data that has no labels. You do not tell the model what to look for or what the right answers are; it finds groups, patterns and structure by itself, and you interpret what it found. It is the natural choice when you want to explore data, such as discovering the topics on your site or the segments in your customer base.

How it works

The dividing line with supervised learning is simple: do you already know the correct answer for at least some of your data? If not, you are in unsupervised territory. The main tasks are:

• Clustering. Grouping similar items, with algorithms such as K-means, DBSCAN, hierarchical clustering and topic models such as LDA and BERTopic.
• Dimensionality reduction. Compressing many features into a few while keeping the important signal, as with PCA and UMAP, often so data can be plotted.
• Anomaly detection. Finding the points that fit no pattern.

Because there are no labels to check against, quality is judged with internal measures (such as silhouette or coherence scores) and, above all, human judgement.

How it’s used in SEO and marketing

• Topic modelling. Discover the themes across a blog or a crawl export.
• Keyword clustering. Group a keyword universe without predefined topics.
• Customer segmentation. Find segments from purchase or engagement behaviour.
• Anomaly detection. Flag unusual traffic spikes, spammy backlinks or odd customer behaviour.
• Exploration before labelling. Cluster first, name the clusters, then use those names as labels for a supervised model.

Example

MLforSEO’s keyword mapping guide compares both approaches on the same keywords. With supervised methods (Sentence-BERT or fuzzy matching), you supply topic labels and each keyword is assigned to one. With BERTopic, unsupervised, you supply only the keywords and the model discovers the topics, including ones you had not thought of.

Limitations and things to watch

• There is no ground truth, so results can look convincing without being useful.
• Output depends heavily on pre-processing, features and settings such as the number of clusters.
• Clusters arrive unnamed; interpreting them takes time and domain knowledge.
• Results can change between runs, so fix random seeds for reproducibility.

Learn it on MLforSEO

• The Characteristics of a Machine Learning Problem
• What Machine Learning Clustering Is, and Where It Fits in SEO
• Different ways to Map Keywords to Topics
• ML Data Characteristics Mapped to SEO and ML Tasks (Guide)
• Introduction to Machine Learning for SEO course