The following are the main points you should have learned from this chapter:

• •

  Relational representations are used when an agent requires models to be given or learned before it knows which entities it will encounter, or when the data includes identifiers, such as part numbers and booking numbers.

• •

  Collaborative filtering and other embedding-based methods can be used to make predictions about instances of relations from other instances by inventing latent properties.

• •

  Plate models allow for the specification of probabilistic models before the entities are known.

• •

  Many of the probabilistic representations in earlier chapters can be made relational by including universally quantified logical variables and parameter sharing.

• •

  The recommendation systems at the heart of many large corporations are optimizing engagement, which leads to increased polarization.