Abstract:
Community detection is crucial for uncovering cohesive substructures within com-
plex systems. These communities provide insights into clusters of interconnected enti-
ties, which can be particularly valuable in various domains such as social network anal-
ysis, information retrieval, and bibliometrics. In this study, we propose a taxonomy
of community detection methods based on graph autoencoders (GAEs), categorizing
them into simple encoder and dual encoder models. We conduct a comparative analy-
sis of these two categories, focusing on the type of encoder architecture and assessing
their performance on real networks. For a more precise evaluation, we use NMI, ARI,
and F1-measure as evaluation metrics. Additionally, we examine the running time
efficiency of each model based on epochs. The findings indicate that dual encoder
models, especially those with attention mechanisms, generally exhibit superior per-
formance, particularly in complex datasets, despite higher computational demands.
These results underscore the potential of dual encoder models in advanced network
analysis tasks. Future recommendations include examining more advanced neural net-
work designs and the impact of modeling and data preparation factors on community
detection across various domains.
