Pruning-based graph neural network de-correlation fine tuning method
By introducing pruning and decorrelation fine-tuning methods into graph neural networks, the high demand for computing resources and insufficient flexibility caused by model redundancy is solved, and the reduction of model parameters and the guarantee of downstream task performance is achieved.
CN119740603APending Publication Date: 2025-04-01BEIHANG UNIV
0 Cites 0 Cited by
Patent Information
- Application Number
- CN202411697508.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-26
- Publication Date
- 2025-04-01
AI Technical Summary
Technical Problem
Graph neural networks have model redundancy problems in the self-supervised learning process, resulting in high demand for computing resources and insufficient flexibility in downstream tasks.
Method used
A pruning-based fine-tuning method for decorrelation of graph neural networks is proposed, which fine-tunes are fine-tuned by random removal of neurons and combining decorrelation methods to optimize model performance and improve computational efficiency.
Benefits of technology
It effectively reduces the number of parameters of the model, ensures the performance of downstream tasks, and provides a more feasible solution for the practical application of graph neural networks.
✦ Generated by Eureka AI based on patent content.
Abstract
The invention provides a graph neural network de-correlation fine tuning method based on pruning. Comprising two parts: a model simplification module and a model decorrelation module; the simplification module is used for reducing model redundancy in a self-supervised pre-training GNN by randomly removing redundant neurons so as to obtain a simplified graph neural network; the model decorrelation module is applied to a fine tuning stage, learns a decorrelation weight used for classifying loss, adjusts weights of fs and R by minimizing a reweighted partial cross covariance matrix, encourages independence of node embedding in different dimensions, so as to reduce redundancy between embedding representations and enable the embedding representations to have more information amount; for graph data, the graph data is input into a simplified graph neural network to obtain embedded representations of nodes, and a label of each node is predicted in combination with an additional prediction layer R.
Need to check novelty before this filing date? Find Prior Art