Incremental Inductive Dynamic Network Community Detection.

Ling Wu,Jiangming Zhuang,Kun Guo

DOI: https://doi.org/10.1007/978-981-99-9637-7_7

2023-01-01

Abstract:In order to address the problem of reconstruction and retraining time overhead in representation learning processing dynamic networks, this paper proposes an incremental inductive dynamic network community detection algorithm (IINDCD). First, the algorithm uses an attention mechanism to capture node neighborhood information and learn node representations by neighborhood aggregation induction while enhancing low-order structural representations. Second, the design uses random walking to capture high-order information and use it to construct node initial features for input into the attentional autoencoder, which effectively fuses high- and low-order structural features. Finally, the algorithm introduces the ideas of incremental update and model reuse for dynamic representation learning, constructs incremental node sets for updating the model, reduces training overhead, and quickly obtains node representation vectors for new moments of the network, then completing dynamic network community detection. IINDCD runs without reconstruction and with low retraining overhead.