Learning Permutation-Invariant Embeddings for Description Logic Concepts

Concept learning deals with description logic concepts from a background knowledge and input examples. The goal is to learn concept that covers all positive examples, while not covering any negative This non-trivial task often formulated as search problem within an infinite quasi-ordered space. Although state-of-the-art models have been successfully applied tackle this problem, their large-scale applications severely hindered due excessive exploration incurring impractical runtimes. Here, we propose remedy for limitation. We reformulate the multi-label classification neural embedding model (NERO) learns permutation-invariant embeddings sets of examples tailored towards predicting $$F_1$$ scores pre-selected concepts. By ranking such in descending order predicted scores, possible can be detected few retrieval operations, i.e., no exploration. Importantly, top-ranked used start procedure symbolic multiple advantageous regions space, rather than starting it most general $$\top $$ . Our experiments on 5 benchmark datasets 770 problems firmly suggest NERO significantly (p-value $$<1\%$$ ) outperforms terms score, number explored concepts, total runtime. provide open-source implementation our approach ( https://github.com/dice-group/Nero ).