Meta-neighbor Aggregated Graph Attention Network for Heterogeneous Graph Representation

• Published in: 2021 IEEE/CIC International Conference on Communications in China (ICCC) pp. 248 - 253
• Main Authors: Zhang, Jincheng, Gan, Xiaoying
• Format: Conference Proceeding
• Language: English
• Published: IEEE 28.07.2021
• Subjects: Aggregates; Fuses; graph neural network; graph representation; heterogeneous graph; Noise measurement; Recommender systems; Semantics
• DOI: 10.1109/ICCC52777.2021.9580337

Graph representation has been widely adopted in many real-world applications, but it's still a challenge for us to represent the heterogeneous information network (HIN). To fuse the rich semantic information, current works usually employ metapaths to generate low dimensional representation of HIN. However, their methods either ignore the node content information, discard intermediate nodes along the metapaths, unaware of distant neighbors or aggregate some noisy nodes. We propose Meta-neighbor Aggregated Graph Attention Network (MNGAT) to deal with these problems and improve the performance of heterogeneous graph embedding. Firstly, our model utilizes multiple metapaths to generate the candidate meta-neighbor set and chooses the top k most frequently appeared neighbors for each target node. Secondly, we use the attention mechanism to assign different weights for each meta-neighbor and leverage the skip connection method to maintain the previous layers' information. Experiments on two real-world datasets demonstrate that our model achieves a better performance over the baselines.