Adaptive Order Aggregator and Extractor Graph Neural Network

Graph Neural Networks (GNNs) have been successful in modeling homophily graphs, but their performance is limited when applied to heterophily graphs. Some methods based on the LSTM can mitigate this issue while they artificially fix the input sequence which is reversed from the collection order of the path. However, our empirical experiments find that different graphs require different input sequences for optimal accuracy. To address these problems, we propose a novel Adaptive Order Aggregator and Extractor Graph Neural Network (AOAENet) to learn the preferences of different order neighbors adaptively. Specifically, we design a dual-input Order Extractor (OE) module using the GRU. We first use the Order Separator (OS) operation to separately obtain each order information, as the current state of OE. Then we propose an Order Aggregator (OA) mechanism to adaptively aggregate each order information, as the hidden state of OE. By iterative usage of OE and OA, we build a deep network to comprehensively extract and adaptively aggregate crucial latent information from each order. Experiments show that AOAENet achieves superior performance on both homophily and heterophily graphs.