Global multivariate model learning from hierarchically correlated data

TL;DR

This paper develops a Bayesian method to accurately infer multivariate Gaussian models from hierarchically correlated data, effectively disentangling intrinsic signals from historical correlations, and significantly improving inference accuracy.

Contribution

It introduces a Bayesian framework for multivariate model learning that accounts for hierarchical correlations, reducing bias and increasing effective sample size.

Findings

Improved accuracy in inferred equilibrium distributions.

Achieves a two- to fourfold increase in effective sample size.

Effectively disentangles intrinsic signals from hierarchical correlations.

Abstract

Inverse statistical physics aims at inferring models compatible with a set of empirical averages estimated from a high-dimensional dataset of independently distributed equilibrium configurations of a given system. However, in several applications such as biology, data result from stochastic evolutionary processes, and configurations are related through a hierarchical structure, typically represented by a tree, and therefore not independent. In turn, empirical averages of observables superpose intrinsic signal related to the equilibrium distribution of the studied system and spurious historical (or phylogenetic) signal resulting from the structure underlying the data-generating process. The naive application of inverse statistical physics techniques therefore leads to systematic biases and an effective reduction of the sample size. To advance on the currently open task of extracting intrinsic signals from correlated data, we study a system described by a multivariate Ornstein-Uhlenbeck process defined on a finite tree. Using a Bayesian framework, we can disentangle covariances in the data corresponding to their multivariate Gaussian equilibrium distribution from those resulting from the historical correlations. Our approach leads to a clear gain in accuracy in the inferred equilibrium distribution, which corresponds to an effective two- to fourfold increase in sample size.