Over-Parameterized Neural Models based on Graph Random Features for fast and accurate graph classification
Lists
Only you see this
Abstract
Recent works have demonstrated the feasibility of fast and accurate time series classification methods based on randomized convolutional kernels. In the context of graph-structured data, most randomized graph neural networks follow the Echo State Network paradigm, where either single layers or the entire network exhibit some form of recurrence. This paper explores a simple form of randomized graph neural network by implementing a graph convolutional network with untrained weights. Node representations are then aggregated through global pooling operators to obtain an untrained graph-level representation. Since no training is involved, computing such representations is extremely fast. A fast linear classifier is subsequently applied to the resulting representations. We show that this simple approach achieves competitive predictive performance while being highly efficient at both training and inference time. In addition, we investigate when over-parameterization, namely generating more features than required for interpolation, can be beneficial for generalization. Using the algorithmic stability framework and empirical evidence from benchmark graph datasets, we provide insights into the role of over-parameterization in this setting.
BibTeX
@inproceedings{navarin2023over,
author = {Nicolò Navarin and Luca Pasa and Claudio Gallicchio and Luca Oneto and Alessandro Sperduti and Nicolò Navarin},
title = {{Over-Parameterized Neural Models based on Graph Random Features for fast and accurate graph classification}},
booktitle = {Zenodo},
year = {2023},
doi = {10.5281/zenodo.18457854},
}Loading related content...