An Information Entropy based Inference Acceleration Method for Graph Neural Networks

• 1、School of Computer Science and Technology, Beijing 100044

Abstract：Due to the exponential growth of graph data and the large number of model parameters, Graph Neural Networks (GNNs) often face high computational costs, which significantly limit their practical application. Recent research has focused on applying the Lottery Ticket Hypothesis (LTH) for the sparsification of GNNs, with the aim of reducing inference costs without compromising performance. This approach, achieved by streamlining the graph structure and model parameters, is designed to enhance the computational efficiency and practicality of GNNs. However, LTH-based methods have two main drawbacks: 1) they require exhaustive iterative training of dense models, leading to substantial training computational costs; and 2) they employ unstructured pruning of model parameters, which currently fails to achieve satisfactory practical acceleration. To overcome these limitations, this paper proposes an information entropy based pruning method for Graph Neural Networks, termed EBP. This method identifies and removes the least impactful edges and model parameter channels for final predictions, accelerating the model while ensuring that the pruning process does not adversely affect model performance. Experimental results show that this pruning method can achieve a good balance between model performance and computational efficiency. The conclusion demonstrates that the entropy-based pruning strategy provides an effective means of accelerating GNN models, facilitating the improvement of GNNs\' feasibility and efficiency in practical applications.

Keywords： artificial intelligence graph neural network graph sparsification model pruning