Unsupervised Learning

Unsupervised learning finds structure in data that carries no labels. Nobody tells the model what the right answer is; it is asked to describe the shape of the data instead — which points resemble each other, which are unlike everything else, what smaller set of dimensions captures most of the variation. It exists because labels are expensive and unlabelled data is nearly free, so a great deal of useful work can be done on what you already have. The main techniques map onto recognisable business jobs. Clustering groups similar records, which is how behavioural customer segments get discovered rather than assumed. Dimensionality reduction compresses many correlated fields into a few informative ones, useful both for visualisation and as a preprocessing step. Anomaly detection learns what normal looks like and flags what doesn't, which underpins fraud screening and infrastructure alerting. Embeddings — the representation behind semantic search and retrieval — come from objectives that need no human labels at all, which is why they scaled so far. The catch is evaluation, and it is a genuine one. With no ground truth there is no accuracy to report, so "is this good?" becomes a judgement call. A clustering algorithm asked for five clusters will cheerfully return five whether or not five is a meaningful number, and the segments it produces may be statistically tidy while corresponding to nothing a marketer can act on. Internal metrics like silhouette scores measure geometric tidiness, not usefulness. The honest test is downstream: does treating these segments differently change a business outcome, does this anomaly flag correlate with real incidents. For a SaaS team the pragmatic pattern is to treat unsupervised results as hypotheses rather than answers. Cluster your accounts, then have someone who knows the customers look at the clusters and name them — if they cannot, the clustering has not found anything real. Used that way it is a genuinely cheap discovery tool. Used as an oracle it produces confident nonsense, and because there is no accuracy number to contradict it, that nonsense can survive a long time.