Statistical Relational Learning

Statistical relational learning (SRL) is a subfield of artificial intelligence that combines two worlds: first-order logic, to represent the relational structure between entities, and probabilistic models, to handle uncertainty. It thus unites reasoning about relationships with the tools of statistics.

The problem it solves

Most machine learning assumes that data are independent and come in a flat table: one row per example, with no connections among them. But a great deal of real data is relational: the people in a social network are connected, proteins interact, the entities of a knowledge graph link to one another. In those cases, the relationships are the information. SRL models precisely that interconnected structure, without flattening it.

Its formalisms

There are several reference models. Markov logic networks (Richardson and Domingos, 2006) attach a weight to each logical formula, turning a set of rules with exceptions into a probabilistic model. Probabilistic relational models (Friedman, Getoor, Koller and Pfeffer, 1999) extend Bayesian networks to relational data. And languages such as ProbLog add probabilities to logic programming in the style of Prolog.

At the intersection of two fields

SRL lives, by design, at a crossroads. It takes the logical part from inductive logic programming and the statistical part from probabilistic graphical models, and fuses them. That is why it is considered a close relative of neuro-symbolic artificial intelligence, which pursues a kindred goal: to reconcile learning with structured reasoning.

What it is for

Its typical tasks are those of networked data: link prediction (guessing which connections are missing or will appear), collective classification (classifying entities by drawing on those in their surroundings) and entity resolution. It is applied to knowledge graphs, information extraction and bioinformatics, where it helps infer unknown relationships between biological entities.