A Complete Decomposition of KL Error using Refined Information and Mode Interaction Selection

TL;DR

This paper introduces a new information geometric approach to decompose KL error in log-linear models, emphasizing higher-order interactions, and proposes an algorithm for sparse mode selection that improves data efficiency.

Contribution

It presents a complete KL error decomposition focusing on higher-order modes and develops MAHGenTa, a novel algorithm for sparse interaction selection in energy-based models.

Findings

MAHGenTa effectively maximizes log-likelihood on synthetic and real datasets.

The approach improves data efficiency by selecting relevant mode interactions.

The method adapts well to both generative and discriminative tasks.

Abstract

The log-linear model has received a significant amount of theoretical attention in previous decades and remains the fundamental tool used for learning probability distributions over discrete variables. Despite its large popularity in statistical mechanics and high-dimensional statistics, the majority of related energy-based models only focus on the two-variable relationships, such as Boltzmann machines and Markov graphical models. Although these approaches have easier-to-solve structure learning problems and easier-to-optimize parametric distributions, they often ignore the rich structure which exists in the higher-order interactions between different variables. Using more recent tools from the field of information geometry, we revisit the classical formulation of the log-linear model with a focus on higher-order mode interactions, going beyond the 1-body modes of independent distributions and the 2-body modes of Boltzmann distributions. This perspective allows us to define a complete decomposition of the KL error. This then motivates the formulation of a sparse selection problem over the set of possible mode interactions. In the same way as sparse graph selection allows for better generalization, we find that our learned distributions are able to more efficiently use the finite amount of data which is available in practice. We develop an algorithm called MAHGenTa which leverages a novel Monte-Carlo sampling technique for energy-based models alongside a greedy heuristic for incorporating statistical robustness. On both synthetic and real-world datasets, we demonstrate our algorithm's effectiveness in maximizing the log-likelihood for the generative task and also the ease of adaptability to the discriminative task of classification.