Box embeddings for extending ontologies: a data-driven and interpretable approach.

Area of Science:

• Artificial Intelligence
• Cheminformatics
• Bioinformatics

Background:

• Deep learning models lack transparency, hindering symbolic knowledge extraction.
• Interpretable AI is crucial for understanding complex model outputs.
• Multi-label classification tasks often involve inherent hierarchical label structures.

Purpose of the Study:

• To develop a method for deriving symbolic knowledge from deep learning models.
• To enforce a taxonomical structure on model outputs for enhanced interpretability.
• To represent implicit logical relationships in multi-label datasets using geometric embeddings.

Main Methods:

• Utilized box-shaped embeddings of ontology classes in vector space.
• Enforced a taxonomical structure on model outputs during training.
• Assessed model performance by approximating subclass relations in the ChEBI ontology.

Main Results:

• The model successfully captures implicit hierarchical relationships among labels.
• Ensured consistency with the underlying ontological conceptualization.
• Achieved state-of-the-art performance in multi-label classification tasks.

Conclusions:

• The proposed approach enables interpretable outputs in chemical classification.
• Geometric representation of molecules and classes facilitates understanding of logical relationships.
• Implicit hierarchies are learned without explicit taxonomy during training.