Explore Alternative Graph Pooling Approaches

[aditya0by0]

Description:
Currently, graph-level embeddings are obtained using simple sum/mean pooling over node embeddings. While effective, this approach may not capture more complex structural information. We should explore alternative graph pooling methods.

Two main alternatives:

1. Set-based pooling with LSTMs + Attention:
   Inspired by Vinyals et al. (2015), this approach iteratively applies attention-based aggregation guided by an LSTM. This allows the pooling process to selectively weight node embeddings and capture richer dependencies.

2. Graph Coarsening / Clustering approaches:
   Instead of treating pooling as a pure set problem, graph topology can be exploited by clustering or coarsening nodes into groups before pooling. This preserves structural information and provides more meaningful graph-level embeddings.

References:

• Section 5.5 Graph Pooling of the Book Graph representation learning / William L. Hamilton, McGill University

[aditya0by0]

https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html#pooling-layers

[aditya0by0]

Section 9.3 Pooling layers: Learning graph-level outputs from node-level outputs from Graph Neural Networks: Foundations, Frontiers, and Applications