Graph Reasoning Networks

TL;DR

Graph Reasoning Networks (GRNs) tackle the limited high-level reasoning of graph neural networks by integrating a semi-learnable encoder with a differentiable MAX-SAT reasoning module (SatNet) trained via an SDP relaxation. The encoder fuses fixed topology-based features with learned representations to produce a vector r in [0,1]^d that the reasoning module uses to derive a discrete class. The approach enables end-to-end gradient-based optimization and can learn rules that standard GNNs miss on synthetic data, while achieving competitive results on real-world datasets with and without node features. This demonstrates the value of combining fixed structural cues with learned representations for graph classification and reasoning tasks.

Abstract

Graph neural networks (GNNs) are the predominant approach for graph-based machine learning. While neural networks have shown great performance at learning useful representations, they are often criticized for their limited high-level reasoning abilities. In this work, we present Graph Reasoning Networks (GRNs), a novel approach to combine the strengths of fixed and learned graph representations and a reasoning module based on a differentiable satisfiability solver. While results on real-world datasets show comparable performance to GNN, experiments on synthetic datasets demonstrate the potential of the newly proposed method.

Figures (1)

• Figure 1: Combining topological and learned node features