A rough guide to the Aldous-Hoover representation theorem for exchangeable arrays
probabilityexchangeabilityrepresentation-theoremmachine-learning
Abstraction: Representation theorem for row-and-column exchangeable random arrays
Key points:
- A sequence is exchangeable if its distribution is permutation-invariant; de Finetti's theorem shows such sequences are conditionally i.i.d. given a latent variable alpha
- David Aldous (1981) extended this to 2D arrays X_ij invariant under row and column permutations (jointly exchangeable / RCE arrays)
- The Aldous-Hoover theorem: any RCE array can be written as X_ij = f(alpha, xi_i, eta_j, lambda_ij) for i.i.d. uniform random variables and a measurable function f
- Applications include modelling networks, graphs, and collaborative filtering where row/column ordering is arbitrary
- Talk by Dr Daniel Roy at Cambridge CBL, May 2012, part of the Machine Learning Reading Group series
Connections: Daniel Roy · Exchangeability · De Finetti Theorem · Random Arrays