Unsupervised Learning

Study guides: Exam SRM, Exam PA, Exam MAS-II

Unsupervised learning looks for structure in predictors X when there is no response Y to predict. With no labels there is no test error to minimize, so results are judged by interpretability and stability rather than by accuracy — the defining contrast with Supervised Learning.

\text{given } X_1, \dots, X_n \in \mathbb{R}^{p}, \text{ find structure — no } Y

• The two MAS-II syllabus methods: Principal Components Analysis (find a few directions that carry most of the variance) and Clustering (find groups of similar observations)
• No objective scorecard. There is no held-out label, so Cross-Validation cannot arbitrate — a Scree Plot elbow or a Dendrogram cut is a judgement call
• Results are scale-sensitive: both PCA and distance-based clustering must have their variables standardized when units differ
• Common actuarial uses: compressing a long list of highly correlated rating variables into a few uncorrelated score variables (principal components), grouping territories or classes, and exploratory data analysis before a Generalized Linear Model
• Often a preprocessing step — principal components fed into a supervised model as predictors

Reading an Unsupervised Result PCA on 12 telematics variables gives a first component with large positive loadings on hard-braking rate, cornering force and night-driving share. What has the analysis found, and what has it not?

Answer It has found that those variables move together — one latent "aggressive driving" direction accounts for much of the spread across drivers, so 12 variables can be summarized by fewer. It has not found that this direction predicts claims. That requires a supervised fit with a claims response; a high-variance direction and a predictive one are different things.