Unsupervised graph neural networks with recurrent features for solving combinatorial optimization problems

Daria Pugacheva, Yuriy Zotov, Andrei Ermakov, Igor Lyskov

Primary Area: learning on graphs and other geometries & topologies

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: graph neural networks, combinatorial optimization, recurrent neural networks, maximum cut problem, graph coloring problem

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

In recent years, graph neural networks (GNNs) have gained considerable attention as a promising approach to tackle combinatorial optimization problems.
We introduce a novel algorithm, dubbed QRF-GNN in the following, that leverages the power of GNNs to efficiently solve combinatorial problems which have quadratic unconstrained binary optimization (QUBO) formulation.
It relies on unsupervised learning and minimizes the loss function derived from QUBO relaxation.
The key components of the architecture are the recurrent use of intermediate GNN predictions, parallel convolutional layers and combination of artificial node features as input.
The performance of the algorithm was evaluated on benchmark datasets for maximum cut and graph coloring problems.
Results of experiments show that QRF-GNN surpasses existing graph neural network based approaches and is comparable to the state-of-the-art conventional heuristics.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7813