DeltaGNN: Graph Neural Network with Information Flow Control

DeltaGNN proposes a new mechanism based on information flow control, aimed at addressing the problems of over-smoothing and over-compression in graph neural networks. This mechanism utilizes information flow scores for optimization, introduces only linear computational overhead, and is supported by theoretical evidence, as well as being scalable and versatile. It can effectively handle long-range and short-range interaction detection. The research team conducted benchmark tests on 10 real-world datasets with different sizes, topologies, densities, and proportions, and the results showed that DeltaGNN performed well at a limited computational complexity. The relevant implementation code is available on the GitHub repository.