[Paper Review] Neural Bipartite Matching

Dobrik Georgiev, Píetro Lió|arXiv (Cornell University)|May 22, 2020

Advanced Graph Neural Networks7 references7 citations

TL;DR

This paper proposes a graph neural network (GNN)-based neural execution framework that learns to perform the Ford-Fulkerson algorithm for maximum bipartite matching by reducing it to a maximum flow problem. Using a single GNN to generate features and iteratively apply augmenting path finding, bottleneck detection, and residual capacity updates, the model achieves near-perfect generalization, attaining optimal matching in 99.8% of test cases across diverse graph scales and edge probabilities.

ABSTRACT

Graph neural networks (GNNs) have found application for learning in the space of algorithms. However, the algorithms chosen by existing research (sorting, Breadth-First search, shortest path finding, etc.) usually align perfectly with a standard GNN architecture. This report describes how neural execution is applied to a complex algorithm, such as finding maximum bipartite matching by reducing it to a flow problem and using Ford-Fulkerson to find the maximum flow. This is achieved via neural execution based only on features generated from a single GNN. The evaluation shows strongly generalising results with the network achieving optimal matching almost 100% of the time.

Motivation & Objective

• To investigate whether graph neural networks (GNNs) can learn to execute complex, multi-step algorithms like Ford-Fulkerson, which involves multiple composable subroutines.
• To evaluate whether GNNs can respect algorithmic invariants, such as flow conservation and capacity constraints, during iterative execution.
• To design a unified neural execution framework that learns all subroutines (path finding, bottleneck detection, augmentation) simultaneously using a single GNN architecture.
• To assess strong generalization performance of the model on unseen graph sizes and edge probabilities beyond the training distribution.

Proposed method

• The model uses an encode-process-decode framework with a GNN processor that operates on residual graphs at each step of the Ford-Fulkerson algorithm.
• Node and edge features are embedded using learnable vectors for each bit position of 8-bit binary capacities, enabling discrete input processing.
• The processor network computes node-level latent features via message passing, and a decoder network predicts augmenting paths and bottleneck capacities.
• Termination is determined by a learned termination network that applies a sigmoid to the global mean of node representations, halting when confidence exceeds 0.5.
• The model is trained end-to-end using supervised signal from ground-truth algorithm steps, with early stopping based on validation accuracy.
• Two GNN architectures are evaluated: MPNN with max aggregation and PNA with standard deviation aggregator removed to prevent overfitting.

Experimental results

Research questions

• RQ1 Can a single GNN architecture learn to execute the multi-step Ford-Fulkerson algorithm for maximum bipartite matching?
• RQ2 Does the model generalize strongly to graph sizes and edge probabilities not seen during training?
• RQ3 Can the GNN learn and preserve algorithmic invariants such as flow conservation and capacity constraints across multiple iterations?
• RQ4 How does the performance of the model compare when subroutines are learned jointly versus individually?
• RQ5 What is the impact of architectural choices like PNA vs. MPNN and removal of the standard deviation aggregator on learning stability and accuracy?

Key findings

• The model achieved 99.8% accuracy in finding the maximum bipartite matching on unseen graphs of size 8, 16, 32, and 64, demonstrating strong generalization.
• With a threshold of 5 for path termination, the model achieved 100% accuracy on all test scales, indicating near-perfect execution of the full algorithm.
• The PNA-based model outperformed MPNN in final accuracy but required more training steps and extra data for the BFS subroutine, showing slower convergence.
• Even with threshold-based termination (e.g., t=1), the model achieved 90% accuracy on 2× scale graphs, indicating robust path-finding capability.
• The model maintained over 99.73% average accuracy across all test sets with varying edge probabilities (1/5, 1/2, 3/4), confirming strong generalization beyond training distribution.
• Ablation studies showed that bottleneck detection and augmentation subroutines had minimal impact on overall accuracy, suggesting path finding is the primary learning bottleneck.