Unsupervised learning is what happens when an AI system is given a large set of data with no correct answers attached at all, and asked to find whatever structure is genuinely there — grouping similar records together, or noticing which features tend to occur together — without anyone telling it in advance what the groups or patterns should be.
The United States’ National Institute of Standards and Technology (a US, not Canadian, source) describes the method plainly in its adversarial-machine-learning taxonomy as the learning paradigm “which trains models using unlabeled data at training time”, in contrast to supervised learning’s reliance on a labelled answer key.
No Canadian statute defines the term. Because unsupervised learning often runs on whatever data is already on hand — there is no labelling step to force a decision about what data is actually needed — the more relevant Canadian anchor is the federal, provincial and territorial privacy regulators’ principle of necessity and proportionality, which instructs organizations: “Use anonymized, synthetic, or de-identified data rather than personal information” where the latter is not required for the purpose.
A retailer with years of purchase history, but no predefined customer categories, can hand that history to a clustering algorithm and let it group customers by behaviour on its own — the algorithm decides where the natural groupings fall, rather than being told in advance which customer counts as “high-value.” That is unsupervised learning at work: no labelled answer exists anywhere in the input data for the algorithm to check itself against.
The same family of technique appears on the defensive side of security research: NIST’s taxonomy notes that unsupervised clustering models are themselves used for tasks like flagging unusual network traffic, precisely because they can surface a pattern nobody thought to label in advance — and, for the same reason, can also be manipulated by anyone who can quietly shift what the underlying data looks like.
See also: supervised learning, reinforcement learning, deep learning.
Segmenting or profiling customer data with an unsupervised model raises the same personal-information questions as any other use of the data — custom-ai-solutions covers how that gets scoped into a build from the outset.
