What is unsupervised learning?
Unsupervised learning is machine learning on data that has no labels. Nobody tells the model the right answer for each record, so instead of predicting an outcome it looks for structure in the inputs themselves. The result is a grouping, a ranking of how unusual each record is, or a simpler representation of the data.
Because no labels are needed, unsupervised methods can run on data a business already holds, such as transaction histories or sensor logs. The trade-off is that the output needs interpretation. A clustering algorithm will return five groups of customers, and a person has to decide whether those groups mean anything.
Common unsupervised learning techniques
- Clustering groups similar records together. The k-means algorithm and Gaussian mixture models are the usual starting points, and the number of groups is a choice the analyst makes.
- Dimensionality reduction compresses many variables into a few while keeping most of the information. Principal component analysis and autoencoders are common tools.
- Anomaly detection learns what normal looks like and flags records that deviate from it, which helps when examples of the problem are rare or unlabeled.
- Association rule mining finds items that occur together, such as products bought in the same basket or faults logged in the same maintenance visit.
Unsupervised learning in practice
A specialty retailer with a loyalty program sends the same promotion to everyone. Purchase frequency, basket size, category mix, and time since last visit are computed for each member, and a clustering algorithm separates the base into distinct groups. One group buys rarely but spends heavily on premium lines, another visits weekly for consumables. Marketing designs a different offer for each, and the groups become labels that a supervised model can later use. A freight broker applies anomaly detection to similar effect, flagging shipments whose weight, route, and cost combination looks unlike anything seen before.
What to watch for
- Results depend heavily on how the features are scaled and chosen, so two analysts can get different clusters from the same data.
- There is no test score in the usual sense. Validate the output by checking whether the groups or anomalies lead to decisions that hold up.
- Clusters drift as behavior changes, so schedule a periodic refresh and compare the new groups with the old ones.
- Treat unsupervised output as a source of hypotheses that supervised learning or a human review then confirms.