43  Unsupervised Methods

← Back to: Classical ML Overview

TipTL;DR

Without labels, the goal shifts from prediction to finding structure. k-means (MacQueen, 1967) groups points into clusters by proximity; PCA (Pearson, 1901) finds the orthogonal directions of greatest variance for linear dimensionality reduction. Both are the classical ancestors of the learned representations that neural networks discover automatically.

43.1 k-means: clustering by proximity

Partition \(n\) points into \(k\) clusters, each represented by a centroid, by minimizing within-cluster squared distance:

\[ \min_{\{C_j\}}\ \sum_{j=1}^{k}\sum_{x\in C_j}\|x - \mu_j\|^2. \]

Lloyd’s algorithm alternates two cheap steps until stable: (1) assign each point to its nearest centroid; (2) recompute each centroid as its cluster’s mean. It is fast and ubiquitous, but assumes roughly spherical, equal-size clusters and needs \(k\) chosen in advance — and the result depends on initialization (k-means++ mitigates this).

43.2 PCA: variance-maximizing directions

PCA finds an orthogonal basis ordered by how much variance each direction captures. The principal components are the top eigenvectors of the covariance matrix \(\Sigma = \tfrac{1}{n}X^\top X\) (equivalently, the SVD of centered \(X\)):

\[ \Sigma v_i = \lambda_i v_i, \qquad \lambda_1 \ge \lambda_2 \ge \cdots \]

Projecting onto the top few components compresses the data while keeping most of its variance — useful for visualization, denoising, and preprocessing. PCA is a linear bottleneck; the autoencoder generalizes it to a non-linear one, and the geometry it studies (directions in a vector space carrying meaning) is the same geometry that makes word embeddings work.

43.3 Timeline context

These methods are the conceptual bridge to representation learning. PCA’s “meaningful directions in a learned space” became the distributed representations of Hinton’s work and then embeddings; clustering intuitions reappear in how trained embedding spaces organize semantically. Neural nets did not replace these — they learn the representation end-to-end rather than fixing it a priori, and k-means/PCA remain the go-to tools for exploratory analysis and compression.

NoteKey takeaway

Unsupervised classical methods extract structure without labels: k-means by grouping, PCA by variance-ranked directions. They are the linear, fixed-form ancestors of the non-linear, learned representations at the heart of the neural era — and still the first tools to reach for in exploratory analysis.

← Back to: Classical ML Overview