Explaining Latent-based models for Link Prediction in Knowledge Graphs

Défense de Monsieur Guillaume Latour

Rejoindre la défense

More and more often ML models are criticised for their lack of interpretability.
One must be able to understand the decision process of the model that led to
the refusal of its mortgage, the diagnosis of a disease, or any legal advice.
The ability to provide an explanation for a prediction is crucial and has been
on the spotlight for a moment now.
Link prediction is an interesting task among the knowledge graph realm due
to its various applications, e.g. user recommendation, fact checking, etc.
As far as we know, the methods providing the best results for link prediction
are based on embeddings, and therefore are not intrinsically comprehensible
by a human.
This work proposes a post-hoc interpretability procedure based on rule mining
that retrieves some insights about the models’ motivations for the provided
predictions.
Keywords: knowledge graph embeddings, link prediction, explicability, rule
mining.

Télecharger : vCal