Understanding Attention and Generalization in Graph Neural Networks
This work addresses the problem of understanding and improving attention mechanisms in GNNs for researchers and practitioners, but it is incremental as it builds on existing insights and proposes alternative training methods.
The paper investigates the role of attention in graph neural networks (GNNs) and finds that its effectiveness varies, with negligible or harmful effects under typical conditions but exceptional gains of over 60% in some classification tasks under specific conditions, which are challenging to achieve in practice.
We aim to better understand attention over nodes in graph neural networks (GNNs) and identify factors influencing its effectiveness. We particularly focus on the ability of attention GNNs to generalize to larger, more complex or noisy graphs. Motivated by insights from the work on Graph Isomorphism Networks, we design simple graph reasoning tasks that allow us to study attention in a controlled environment. We find that under typical conditions the effect of attention is negligible or even harmful, but under certain conditions it provides an exceptional gain in performance of more than 60% in some of our classification tasks. Satisfying these conditions in practice is challenging and often requires optimal initialization or supervised training of attention. We propose an alternative recipe and train attention in a weakly-supervised fashion that approaches the performance of supervised models, and, compared to unsupervised models, improves results on several synthetic as well as real datasets. Source code and datasets are available at https://github.com/bknyaz/graph_attention_pool.