Chapter 17 Relational Learning and Probabilistic Reasoning

The mind is a neural computer, fitted by natural selection with combinatorial algorithms for causal and probabilistic reasoning about plants, animals, objects, and people.

In a universe with any regularities at all, decisions informed about the past are better than decisions made at random. That has always been true, and we would expect organisms, especially informavores such as humans, to have evolved acute intuitions about probability. The founders of probability, like the founders of logic, assumed they were just formalizing common sense.

– Steven Pinker [1997, pp. 524, 343]

In the machine learning and probabilistic models presented in earlier chapters, the world is made up of features and random variables. As Pinker points out, we generally reason about things. Things are not features or random variables; it doesn’t make sense to talk about the probability of an individual animal, but you could reason about the probability that it is sick, based on its symptoms. This chapter is about how to learn and make probabilistic predictions about things or entities.

The representation dimension has, as its top level, reasoning in terms of individuals (entities) and relations. Reasoning in terms of relations allows for compact representations that can be built independently of the particular entities, but can also be used to learn and reason about each entity. This chapter outlines how feature-based representations as used in learning and probabilistic reasoning can be expanded to deal also with entities and relations. A relational model can benefit from being able to be built before the entities are known and, therefore, before the features are known.

Statistical relational AI or neuro-symbolic AI involves making predictions about relations based on relational data consisting of a relational database, and perhaps metadata. Statistical relational AI is a general term for relational predictions based on data, whereas neuro-symbolic AI involves using neural networks and other embedding-based techniques for predictions. This encompasses:

  • Predicting attributes of entities based on their other attributes and attributes of entities they are related to.

  • Predicting relations based on properties and relations of the entities involved. The simplest case is learning a single binary relation (Section 17.2.1), which is useful for domains like movie recommendations. This is extended to learning triples in Section 17.2.2, and learning more general relations in Section 17.3.

  • Predicting identity, whether descriptions denote the same entity – the descriptions are equal – for example, which citations refer to the same papers, or whether two descriptions refer to the same person (Section 17.4).

  • Predicting existence, whether an entity exists that fits a description, for example whether there is a person in a particular room (Section 17.4).

The set of all entities of a type is called the population. In this chapter, a property where the range is an entity (e.g., a person or a movie) is called a relation and the term property is used when the range is a fixed set, such as Boolean or the reals.