Unsupervised Learning

Read these first

1. 01Gaussian Mixture Models

   A Gaussian mixture describes data as several bell curves blended together, each with its own centre, width and share of the population. Instead of a hard label it gives every point a probability of belonging to each group.

2. 02Hierarchical Agglomerative Clustering

   Start with every point as its own cluster, repeatedly glue the two closest together, and record the whole history as a tree. You get every possible number of clusters at once and choose the granularity afterwards by cutting the tree.

3. 03K-Means Clustering

   The default way to split unlabelled data into k groups, guess k centres, assign every point to its nearest, move each centre to the middle of what it caught, repeat. Fast and useful, but it needs you to pick k and it can land in the wrong answer.

4. 04The Expectation-Maximization Algorithm

   EM solves the chicken-and-egg problem of fitting a model when some labels are missing: guess the labels from the parameters, refit the parameters from the guessed labels, repeat. Each round provably never makes the fit worse.

Then the rest

1. 05Asset Network Graphs and Minimum Spanning Trees
2. 06Choosing the Number of Clusters
3. 07DBSCAN Density Clustering
4. 08Factor Analysis and Latent Variables
5. 09Graph Community Detection
6. 10HDBSCAN
7. 11Independent Component Analysis
8. 12Isolation Forests for Anomaly Detection
9. 13Non-Negative Matrix Factorization
10. 14Sparse PCA
11. 15t-SNE
12. 16UMAP
13. 17Kernel PCA
14. 18Manifold Learning and Isomap
15. 19Spectral Clustering