A Neuro-Symbolic System over Knowledge Graphs for Link Prediction

Tracking #: 3203-4417

This paper is currently under review
Ariam Rivas
Diego Collarana
Maria Torrente
Maria-Esther Vidal

Responsible editor: 
Guest Editors NeSy 2022

Submission type: 
Full Paper
Neuro-Symbolic focuses on integrating symbolic and sub-symbolic systems. The aim is to provide a neural-symbolic implementation of logic, a logical characterization of a neural system, or a hybrid learning system that contributes features of symbolic and sub-symbolic systems. They differ fundamentally in how they represent data and information. Neuro-symbolic systems have recently received significant attention in the scientific communities. However, despite efforts in neural-symbolic integration, symbol processing currently has limited scope and applicability. This work leverages the symbolic system, independent of the application domain, and improves the predictive capability of Knowledge Graph Embeddings (KGE). We tackle the problem of Neuro-Symbolic AI integration, enabling expressive reasoning and robust learning to discover relationships over a knowledge graph. We present a novel approach to integrating Neuro-Symbolic AI systems. Deductive databases implement the symbolic system for an abstract target prediction over a knowledge graph. The symbolic system enhances the predictive capacity of the subsymbolic systems implemented by KGE models. Our approach builds the ego networks of the head and tail of the abstract target prediction, and the symbolic system deduces new relationships enhancing the ego networks. Thus, the subsymbolic systems increase the predictive capacity of the abstract target prediction. As a proof of concept, we have implemented our neuro-symbolic system on top of a KG for lung cancer to predict treatment effectiveness. Our empirical results put the deduction power of deductive databases into perspective; they suggest that enhancing the neighborhoods of the entities on the head or tail of a target prediction can improve the predictive capacity of existing KGE models.
Full PDF Version: 
Under Review