Class Expression Learning with Permutation-Invariant Embeddings

Anonymous

Keywords: Description Logic, Class Expression Learning, Permutation-invariant Representations

TL;DR: We propose a permutation-invariant neural model to learn description logic concept expressions from input example sets and an RDF knowledge base.

Abstract:

Class expression learning deals with learning description logic concepts from an RDF knowledge base and input examples.
The goal is to learn a concept that covers
all positive examples, while not covering any negative examples.
Although state-of-the-art models have been successfully applied to tackle this problem, their large-scale applications have been severely hindered due to their impractical runtimes.
Arguably, this limitation stems from their needs for exploring numerous expressions.
Here, we investigate a remedy for this limitation.
We formulate the class expression learning problem as a multi-label classification problem and we propose a permutation-invariant embedding model (Nero) to reduce the rate of exploration.
For a given learning problem, Nero accurately predicts quality of pre-selected description logic concepts for a given input example sets.
Through ranking concepts in descending order of predicted qualities, the standard search procedure can start in multiple advantageous regions of the quasi-ordered search space.
Our experiments on 5 benchmark datasets with 770 learning problems suggest that using \approach led to significant improvements (p-value <1\%) in the number of explored expressions and the total runtime time.