HopLearn: Leveraging Multi-Hop Neighbors and Learnable Parameters for GNNs with Missing Node Features

Graph neural networks (GNNs) are becoming increasingly central to various applications. A common presumption underpinning their utility is the completeness of node features in graphs. However, this often proves impractical in real-world scenarios where acquiring comprehensive feature information for every node remains a challenge. To address this problem in node classification task, we present a new method called Leveraging Multi-Hop Neighbors and Learnable Parameters (HopLearn) which is simple and easily implementable. In contrast to conventional methods that substitute missing node features with features extracted from multi-hop neighbors, HopLearn not only performs this substitution but also incorporates certain replaced node features as learnable parameters. We evaluate HopLearn using six benchmark datasets across both homophilic and heterophilic graphs, i.e., classes or attributes are often different between adjacent nodes. The results demonstrate that HopLearn exhibits robust efficacy across diverse graph structures. Moreover, HopLearn is compatible with various GNNs and offers rapid processing speeds. Given these attributes, HopLearn is beneficial for applications harnessing GNNs with incomplete graph datasets.