Unsupervised Learning Methods

Article BigQuant

Summary

This overview surveys unsupervised learning through two groups of methods: dimensionality reduction and clustering. For dimensionality reduction, it describes PCA as a way to project high-dimensional features into fewer dimensions using covariance-matrix eigenvectors or singular value decomposition. It also mentions kernel PCA, manifold learning methods such as locally linear embedding, and t-SNE, noting the latter’s use in visualizing complex structures.

For clustering, the document lists hierarchical clustering, K-means, expectation-maximization, and spectral clustering. It sketches K-means as iteratively assigning observations to centers and updating those centers, and EM as alternating expectation and likelihood-maximization steps. The material is a brief historical and conceptual survey attributed to a securities research source; it provides no algorithm comparisons, implementation guidance, trading applications, or empirical results. Its simplified descriptions should therefore be treated as an introduction rather than a practical selection guide.

Key ideas

PCA reduces feature dimensions by retaining directions associated with larger eigenvalues.
Kernel PCA applies PCA after mapping data into a kernel-defined feature space.
t-SNE is presented as a nonlinear method often used to visualize high-dimensional structure.
The clustering methods surveyed include hierarchical clustering, K-means, EM, and spectral clustering.
The overview supplies no comparative evidence for choosing a method in a particular trading task.

Unsupervised Learning Methods for Dimensionality Reduction and Clustering

Summary

Key ideas

Tags