Adaptive Universal Generalized PageRank Graph Neural Network (GPR‑GNN): Overview, Challenges, and Experimental Insights

The talk introduces the paper "Adaptive Universal Generalized PageRank Graph Neural Network" (GPR‑GNN), a recent ICLR contribution that aims to improve graph neural networks (GNNs) for tasks such as node classification, link prediction, and biomedical applications.

It first reviews standard GNN concepts, defining the adjacency matrix \(A\), feature matrix \(X\), degree matrices, and the typical stacking of GNN layers (e.g., GCN, GAT, GraphSAGE). The authors point out two pervasive problems of existing GNNs: (1) limited generality—most models assume homophily and perform poorly on heterophilic graphs, and (2) over‑smoothing—deep stacking quickly drives node representations to a uniform state, degrading performance.

GPR‑GNN addresses these issues with a two‑part architecture: a single‑layer MLP extracts latent node features, followed by \(K\) propagation steps on the graph. Each step produces a representation \(H_k\); a set of learnable GPR weights linearly combines all \(H_k\) to form the final output. This design keeps the parameter count low while allowing deep propagation, thereby mitigating over‑smoothing and enabling the model to adapt to both low‑pass (homophilic) and high‑pass (heterophilic) filtering regimes.

The paper provides theoretical analysis showing that GPR weights correspond to coefficients of a polynomial graph filter. Positive‑only weights implement a low‑pass filter, while alternating sign weights realize a high‑pass filter, explaining how GPR‑GNN can achieve generality across graph types.

Extensive experiments are reported:

On synthetic contextual stochastic block model (cSBM) graphs, GPR‑GNN exhibits symmetric performance when the graph is homophilic (\(\phi>0\)) or heterophilic (\(\phi<0\)), unlike traditional GNNs whose accuracy drops sharply on heterophilic graphs.

On several real‑world datasets, GPR‑GNN consistently outperforms baseline GNNs, confirming its robustness.

Analysis of learned GPR weights shows interpretable patterns: all‑positive weights on homophilic graphs (low‑pass) and alternating‑sign weights on heterophilic graphs (high‑pass).

A study of training dynamics demonstrates that the model automatically reduces the weight of the deepest propagation step, effectively avoiding over‑smoothing and achieving near‑perfect accuracy after training.

The authors conclude that GPR‑GNN offers a general, over‑smoothing‑resistant, and interpretable solution for graph learning, and they suggest future work such as replacing the MLP with more complex networks, learning GPR weights via attention mechanisms, and extending the approach to graph representation learning with pooling layers.