A FAIR COMPARISON OF GRAPH NEURAL NETWORKS FOR GRAPH CLASSIFICATION
A Fair Comparison of Graph Neural Networks for Graph Classification
Federico Errica, Macro Podda, Davide Bacciu, Alessio Micheli
Intro
Evaluate 5 models across 9 benchmarks.
- Provide a fair comparison among GNN architectures. Perform a large number of experiments.
- Explore to what extent current GNN models can effectively exploit graph structure.
- Study the effect of node degrees as a feature in social datasets. Including degrees is beneficial to the performance, and it also has implications in the number of GNN layers needed to reach good results.
Experiments
Baselines 用了两个不同的baselines: chemical 上用的是Molecular Fingerprint。 social 上用的是MLP (node features) + global sum pooling + MLP。 并不包含/利用graph topology。
因此是structure-agnostic:将structure info和node info很好的分离开来。 而且能够发现一旦这个baseline和GNN的差别不大,那就表示GNN not exploiting the structure for graph adequately。 而GNN明显比baselines好的时候,表示topology被很好的利用了。
Results and Discussion
- Table 3:只有一个数据集上GNN比(且都比)baselines好,其余都是baselines更优。
- Table 4:增加node degrees能明显提高performance。