SGKD: A Scalable and Effective Knowledge Distillation Framework for Graph Representation Learning
Document Type
Conference Proceeding
Publication Date
1-1-2022
Abstract
As Graph Neural Networks (GNNs) are widely used in various fields, there is a growing demand for improving their efficiency and scalablity. Knowledge Distillation (KD), a classical methods for model compression and acceleration, has been gradually introduced into the field of graph learning. More recently, it has been shown that, through knowledge distillation, the predictive capability of a well-trained GNN model can be transferred to lightweight and easy-to-deploy MLP models. Such distilled MLPs are able to achieve comparable performance as their corresponding G NN teachers while being significantly more efficient in terms of both space and time. However, the research of KD for graph learning is still in its early stage and there exist several limitations in the existing KD framework. The major issues lie in distilled MLPs lack useful information about the graph structure and logits of teacher are not always reliable. In this paper, we propose a Scalable and effective graph neural network Knowledge Distillation framework (SGKD) to address these issues. Specifically, to include the graph, we use feature propagation as preprocessing to provide MLPs with graph structure-aware features in the original feature space; to address unreliable logits of teacher, we introduce simple yet effective training strategies such as masking and temperature. With these innovations, our framework is able to be more effective while remaining scalable and efficient in training and inference. We conducted comprehensive experiments on eight datasets of different sizes - up to 100 million nodes - under various settings. The results demonstrated that SG KD is able to significantly outperform existing KD methods and even achieve comparable performance with their state-of-the-art GNN teachers.
Identifier
85148431464 (Scopus)
ISBN
[9798350346091]
Publication Title
IEEE International Conference on Data Mining Workshops Icdmw
External Full Text Location
https://doi.org/10.1109/ICDMW58026.2022.00091
e-ISSN
23759259
ISSN
23759232
First Page
666
Last Page
673
Volume
2022-November
Recommended Citation
He, Yufei and Ma, Yao, "SGKD: A Scalable and Effective Knowledge Distillation Framework for Graph Representation Learning" (2022). Faculty Publications. 3306.
https://digitalcommons.njit.edu/fac_pubs/3306
