Existing research methods for analyzing the importance of node features in graph neural networks are primarily based on structural features, making the analysis of each node feature’s importance difficult. To address this issue, a node feature dimension augmentation analysis method for graph neural networks is proposed. Firstly, node features are represented in a high-dimensional space, and a data structure adapted for the sparse characteristics of high-dimensional node feature data is constructed. Then, the computational rules of graph neural networks are extended and defined in high-dimensional space, with high-dimensional node features being calculated accordingly. Finally, the calculation results in high-dimensional space are analyzed, and the weights of each node feature in the graph neural network calculation results are obtained. In experiments, it is found that node features with large weights have the greatest impact on model accuracy, accounting for 62.88%, which verifies that the weights can effectively reflect the importance of input node features.
NAIRV, HINTONG E. Rectified linear units improve restricted boltzmann machines[C]∥Proceedings of the 27th International Conference on International Conference on Machine Learning. Cambridge,USA:Omnipress, 2010:807-814.
[15]
SEN P, NAMATAG, BILGICM,et al. Collective classificationin network data[J]. AI Magazine, 2008,29(3):93-106.
[16]
MCCALLUMA K, NIGAMK, RENNIEJ,et al. Automating the construction of Internet portals with machine learning[J]. Information Retrieval, 2000,3(2):127-163.
[17]
ROSSIR A, AHMEDN K. The network data repository with interactive graph analytics and visualization[C]∥Proceedings of the 29th AAAI Conference on Artificial Intelligence. Washington, USA: AAAI, 2015:4292-4293.