Graph Bernoulli Pooling

Guangbu Liu, Tong Zhang, Xudong Wang, Chuanwei Zhou, Zhen Cui

Keywords: graph classification, variational inference, Bernoulli sampling, mutual information

Abstract:

Graph pooling is crucial for enlarging receptive field and reducing computational cost in deep graph representation learning. In this work, we propose a simple but effective non-deterministic graph pooling method, called graph Bernoulli pooling (BernPool), to facilitate graph feature learning.
In contrast to most graph pooling methods with deterministic modes, we design a probabilistic Bernoulli sampling to reach an expected sampling rate through deducing a variational bound as the constraint. To further mine more useful info, a learnable reference set is introduced to encode nodes into a latent expressive probability space. Hereby the resultant Bernoulli sampling would endeavor to capture salient substructures of the graph while possessing much diversity on sampled nodes due to its non-deterministic manner. Considering the complementarity of node dropping and node clustering, further, we propose a hybrid graph pooling paradigm to combine a compact subgraph (via dropping) and a coarsening graph (via clustering), in order to retain both representative substructures and input graph info. Extensive experiments on multiple public graph classification datasets demonstrate that our BernPool is superior to various graph pooling methods, and achieves state-of-the-art performance. The code is publicly available in an anonymous format at \href{https://anonymous.4open.science/r/BernPool}{https:/github/BernPool}.

Submission Number: 3310