Abstract:Graph Neural Networks (GNNs) have garnered significant attention for their success in learning the representation of homophilic or heterophilic graphs. However, they cannot generalize well to real-world graphs with different levels of homophily. In response, the Possion-Charlier Network (PCNet) \cite{li2024pc}, the previous work, allows graph representation to be learned from heterophily to homophily. Although PCNet alleviates the heterophily issue, there remain some challenges in further improving the efficacy and efficiency. In this paper, we simplify PCNet and enhance its robustness. We first extend the filter order to continuous values and reduce its parameters. Two variants with adaptive neighborhood sizes are implemented. Theoretical analysis shows our model's robustness to graph structure perturbations or adversarial attacks. We validate our approach through semi-supervised learning tasks on various datasets representing both homophilic and heterophilic graphs.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is the challenges faced by graph neural networks (GNNs) when processing real - world graph data with different levels of homophily. Specifically: 1. **Balance between Homophily and Heterophily**: Existing GNNs usually perform well on homophilic graphs (i.e., connected nodes tend to belong to the same class), but not well on heterophilic graphs (i.e., connected nodes tend to belong to different classes). This limits their application in complex real - world graph data. The method proposed in the paper aims to process both homophilic and heterophilic graphs simultaneously, thereby improving the generalization ability of the model. 2. **Robustness of the Model**: Real - world graph data are often affected by measurement errors or other unexpected situations, resulting in structural perturbations or adversarial attacks. Existing GNNs are very sensitive to these perturbations, and their performance will decline significantly. The method proposed in the paper solves this problem by designing filters that are robust to structural perturbations and adversarial attacks. 3. **Parameter Efficiency**: Existing GNNs, such as PCNet, although perform well in processing heterophilic graphs, have too many model parameters, are prone to overfitting, and affect the generalization ability. The method proposed in the paper simplifies PCNet by reducing the number of parameters while maintaining its performance on node classification tasks. ### Main Contributions 1. **Extension of Filter Order**: The paper proposes to extend the filter order to continuous values and reduce the number of parameters, thereby providing more flexible filter design. 2. **Adaptive Neighborhood Size**: Two variants with adaptive neighborhood sizes are implemented, enabling the model to dynamically adjust the neighborhood size according to different nodes. 3. **Robustness Analysis**: A detailed robustness analysis is provided, demonstrating the resistance of the proposed method to graph - structure perturbations and adversarial attacks. 4. **Experimental Verification**: The effectiveness of the proposed method is verified through semi - supervised learning tasks on multiple datasets, showing its advantages in processing homophilic and heterophilic graph data. ### Experimental Results - **Synthetic Graph Experiments**: On synthetic graphs, SPCNet - D performs well in various settings, especially when processing sparse graphs and heterophilic graphs. - **Real - Graph Experiments**: On multiple benchmark datasets, SPCNet - D and SPCNet - L both achieve the best or near - best performance in node classification tasks, especially when processing heterophilic graphs. In conclusion, this paper not only improves the performance of the model in processing homophilic and heterophilic graph data by proposing a simplified PCNet (SPCNet), but also enhances the robustness and parameter efficiency of the model.

Simplified PCNet with Robustness

GCNH: A Simple Method For Representation Learning On Heterophilous Graphs

Universally Robust Graph Neural Networks by Preserving Neighbor Similarity

Exploiting Neighbor Effect: Conv-Agnostic GNN Framework for Graphs with Heterophily

Graph Neural Networks with Heterophily

Neighborhood Homophily-based Graph Convolutional Network

Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs

Hetero$^2$Net: Heterophily-aware Representation Learning on Heterogenerous Graphs

Beyond smoothness: A general optimization framework for graph neural networks with negative Laplacian regularization

Powerful Graph Convolutioal Networks with Adaptive Propagation Mechanism for Homophily and Heterophily

p-Laplacian Based Graph Neural Networks

Addressing Heterogeneity and Heterophily in Graphs: A Heterogeneous Heterophilic Spectral Graph Neural Network

Clarify Confused Nodes via Separated Learning

The Heterophilic Snowflake Hypothesis: Training and Empowering GNNs for Heterophilic Graphs

Heterophilous Distribution Propagation for Graph Neural Networks

Learn from Heterophily: Heterophilous Information-enhanced Graph Neural Network

Seeking Similarities While Removing Differences: Graph Neural Networks Based on Node Correlation.

LSGNN: Towards General Graph Neural Network in Node Classification by Local Similarity

Similarity-navigated Graph Neural Networks for Node Classification

DCGNN: Adaptive deep graph convolution for heterophily graphs

Heterogeneous Graph Propagation Network