Rule Mining in Feature Space

Stefano Teso, Andrea Passerini

Abstract:

Relational embeddings have emerged as an excellent tool for inferring novel facts
from partially observed knowledge bases. Recently, it was shown that some
classes of embeddings can also be exploited to perform a simplified form of rule
mining. By interpreting logical conjunction as a form of composition between re-
lation embeddings, simplified logical theories can be mined directly in the space
of latent representations. In this paper, we present a method to mine full-fledged
logical theories, which are significantly more expressive, by casting the semantics
of the logical operators to the space of the embeddings. In order to extract relevant
rules in the space of relation compositions we borrow sparse reconstruction pro-
cedures from the field of compressed sensing. Our empirical analysis showcases
the advantages of our approach.

TL;DR: We propose an algorithm to discover logical theories from relational embeddings of knowledge bases.

Conflicts: unitn.it

Keywords: Unsupervised Learning