Reference classes and relational learning

This paper studies the connections between relational probabilistic models and reference classes, with specific focus on the ability of these models to generate the correct answers to probabilistic queries. We distinguish between relational models that represent only observed relations and those which additionally represent latent properties of individuals. We show how both types of relational models can be understood in terms of reference classes, and that learning such models correspond to different ways of identifying reference classes. Rather than examining the impact of philosophical issues associated with reference classes on relational learning, we directly assess whether relational models can represent the correct probabilities of a simple generative process for relational data. We show that models with only observed properties and relations can only represent the correct probabilities under restrictive conditions, whilst models that also represent latent properties avoids such restrictions. As such, methods for acquiring latent-property models are an attractive alternatives to traditional ways of identifying reference classes. Our experiments on synthetic as well as real-world domains support the analysis, demonstrating that models with latent relations are significantly more accurate than those without latent relations.

► We show that well-known relational models compose of reference classes.
► Such models are biased when hidden attributes underlie relational data.
► Explicitly accounting for hidden attributes can correct this error.
► The formal argument is validated in synthetic and real-world experiments.