Ming Gu,Zhuonan Zheng,Sheng Zhou,Meihan Liu,Jiawei Chen,Tanyu Qiao,Liangcheng Li,Jiajun Bu
Abstract:Graph Neural Networks (GNNs) have demonstrated remarkable success in various domains, such as transaction and social net-works. However, their application is often hindered by the varyinghomophily levels across different orders of neighboring nodes, ne-cessitating separate model designs for homophilic and heterophilicgraphs. In this paper, we aim to develop a unified framework ca-pable of handling neighborhoods of various orders and homophilylevels. Through theoretical exploration, we identify a previouslyoverlooked architectural aspect in multi-hop learning: the cascadedependency, which leads to asmoothness-generalization <a class="link-external link-http" href="http://dilemma.This" rel="external noopener nofollow">this http URL</a> dilemma significantly affects the learning process, especiallyin the context of high-order neighborhoods and heterophilic <a class="link-external link-http" href="http://graphs.To" rel="external noopener nofollow">this http URL</a> resolve this issue, we propose an Inceptive Graph Neural Net-work (IGNN), a universal message-passing framework that replacesthe cascade dependency with an inceptive architecture. IGNN pro-vides independent representations for each hop, allowing personal-ized generalization capabilities, and captures neighborhood-wiserelationships to select appropriate receptive fields. Extensive ex-periments show that our IGNN outperforms 23 baseline methods,demonstrating superior performance on both homophilic and het-erophilic graphs, while also scaling efficiently to large graphs.
What problem does this paper attempt to address?
### Problems the paper attempts to solve
This paper aims to solve the **smoothness - generalization dilemma** faced by Graph Neural Networks (GNNs) when dealing with different - order neighbors and levels of homophily. Specifically, existing GNN methods usually need to design models for homophilic graphs and heterophilic graphs separately, which limits their versatility and application scope.
#### Background problems
1. **Homophily and heterophily**: Nodes in a graph can be divided into homophilic graphs and heterophilic graphs according to the label similarity of their neighbors. In homophilic graphs, adjacent nodes tend to have similar labels, while in heterophilic graphs, the label differences of adjacent nodes are relatively large.
2. **Influence of multi - order neighbors**: In practical applications, the proportion of homophily in different - order neighbors of a graph may change significantly. For example, in the Cora dataset, first - order neighbors show a high level of homophily, but as the neighbor order increases, homophily drops rapidly; while in the Texas dataset, second - order neighbors show a high level of homophily.
3. **Limitations of existing methods**: Traditional GNN methods rely on first - order neighbor information for aggregation and are difficult to effectively handle the complexity brought by multi - order neighbors, especially in heterophilic graphs.
#### Smoothness - generalization dilemma
Through theoretical analysis, the paper discovers a key problem: in multi - order learning, the traditional cascade dependency will lead to the smoothness - generalization dilemma. Specifically:
- **Smoothness**: Refers to the ability of GNNs to make node representations closer in the network. Especially in homophilic graphs, this smoothness helps to improve the generalization ability.
- **Generalization ability**: Refers to the ability of the model to adapt to changes in different neighborhood distributions. Especially in heterophilic graphs, complex neighborhood distributions require stronger generalization ability.
However, as the number of network layers increases, cascade dependency leads to over - smoothing, thereby weakening the model's generalization ability. This dilemma is particularly evident in high - order neighborhoods and heterophilic graphs.
### Solutions
To solve this dilemma, the paper proposes a new Graph Neural Network framework - **Inceptive Graph Neural Network (IGNN)**. The main features of this framework include:
1. **Separative Neighborhood Transformation (SN)**: Independently learn representations for each - order neighbor, avoid sharing or coupling transformation layers, thereby achieving personalized generalization ability.
2. **Inceptive Neighborhood Aggregation (IN)**: Simultaneously embed multiple different receptive fields, such as different - order neighbors or custom - defined relational neighborhoods, and avoid calculating high - order neighborhood representations based on low - order neighborhoods.
3. **Neighborhood Relationship Learning (NR)**: Learn the correlations between different neighborhoods, capture their commonalities and differences, and further enhance the model's generalization ability.
Through these designs, IGNN can effectively handle different - order neighbors and levels of homophily without relying on specific graph structures, thereby improving the versatility and performance of GNNs.
### Experimental verification
The experimental results show that IGNN significantly outperforms 23 baseline methods on multiple benchmark datasets. It not only performs well on homophilic graphs and heterophilic graphs but also can be efficiently extended to large - scale graph data.
In summary, by revealing and solving the smoothness - generalization dilemma, this paper proposes a general - purpose Graph Neural Network framework IGNN, which significantly improves the performance and applicability of GNNs on various graph structures.