Go to ICLR 2025 Conference homepageGenetic-evolutionary Graph Nerual Networks: A Paradigm for Improved Graph Representation Learning
Message-passing graph neural networks have become the dominant framework for learning over graphs. However, empirical studies continually show that message-passing graph neural networks tend to generate over-smoothed representations for nodes after iteratively applying message passing. This over-smoothing problem is a core issue that limits the representational capacity of message-passing graph neural networks. We argue that the fundamental problem with over-smoothing is a lack of diversity in the generated embeddings, and the problem could be reduced by preserving the embedding diversity in their generation process. To this end, we propose genetic-evolutionary graph neural networks, a new paradigm for graph representation learning inspired by genetic algorithms. We model each layer of a graph neural network as an evolutionary process and develop operations based on crossover and mutation to prevent embeddings from becoming similar to one another, thus enabling the model to generate improved graph representations. The proposed framework is interpretable, as it directly draws inspiration from genetic algorithms for preserving population diversity. We experimentally validate the proposed framework on six benchmark datasets on different tasks. The results show that our method significant advances the performance current graph neural networks, resulting in new state-of-the-art results for graph representation learning on the datasets.