Due to the homophily assumption in graph convolution networks, a common
consensus is that graph neural networks (GNNs) perform well on homophilic
graphs but may fail on heterophilic graphs with many inter-class edges. In this
work, we re-examine the heterophily problem of GNNs and investigate the feature
aggregation of inter-class neighbors. Instead of treating the inter-class edges
as harmful, under some conditions, we argue that they may provide helpful
information for node classification from an entire neighbor's perspective. To
better evaluate whether the neighbor is helpful for the downstream tasks, we
present the concept of the neighbor effect of each node and use the von Neumann
entropy to measure the randomness/identifiability of the neighbor distribution
for each class. Moreover, we propose a Conv-Agnostic GNNs framework (CAGNNs) to
enhance the performance of GNNs on heterophily datasets by learning the
neighbor effect for each node. Specifically, we first decouple the feature of
each node into the discriminative feature for downstream tasks and the
aggregation feature for graph convolution. Then, we propose a shared mixer
module for all layers to adaptively evaluate the neighbor effect of each node
to incorporate the neighbor information. Experiments are performed on nine
well-known benchmark datasets for the node classification task. The results
indicate that our framework is able to improve the average prediction
performance by 9.81%, 25.81%, and 20.61% for GIN, GAT, and GCN, respectively.
Extensive ablation studies and robustness analysis further verify the
effectiveness, robustness, and interpretability of our framework.

本文探讨了 GNNs 的异质问题并研究了跨类别邻居的特征聚合，提出了 CAGNNs 框架以提高节点分类任务的性能。实验结果表明，我们的框架能够分别提高 GIN，GAT 和 GCN 的平均预测准确率 9.81％，25.81％和 20.61％。