Transductive Learning for BI-RADS Knowledge Graph based on Knowledge Tensor Factorization

Abstract

The advantage of Knowledge Graph (KG) can greatly prompt the interpretability of the artificial intelligence diagnosis. For breast ultrasound, the KG can be built through BI-RADS semantic descriptions, and the diagnosis can be achieved by link reconstruction between patients and outcomes. However, the existing KG analysis methods consider only the linked neighbors of the entities and relations during embedding, but not the whole entities and relations in KG, which reduces the link reconstruction power for diagnosis in the case of only a small fraction of labeled patients. In this paper, we present a transductive learning based Knowledge Tensor Factorization (KTF) method, which can effectively represent the KG data through a core tensor of interactions among all entities and relations and their embedding vectors. KTF demonstrates distinct diagnosis performance even if there is only a small fraction of labeled patients. Through experiments of assessments, KTF shows distinct superior performance in diagnosis for KG data of BI-RADS with a small fraction of known outcomes of patients.